VOL. 01 — NO. 09 AHMEDABAD, IN · --:--:--
EN
GITHUB ↗
AI/ML ENGINEER EDITION — 2026

Krish Patel

AI/ML ENGINEER · OCR & DOC AI · COMPUTER VISION · AGENTIC LLMS SECURITY & RELIABILITY A+ verify audit ↓

Python and AI/ML Engineer with hands-on experience in OCR pipelines, Intelligent Document Processing, Computer Vision (YOLOv8), and Agentic LLMs. Building production-grade document extraction, PII masking, and multi-step reasoning systems.

BOOK 30 MINUTES — FREE, NO COMMITMENT
TODAY'S DISPATCH · AHMEDABAD, GUJARAT (IST / UTC+5:30)
STATUS
OPEN TO AI/ML ROLES & PRODUCTION CONTRACTS
ACTIVE FOCUS
SURYA & CHANDRA OCR, BYT5 NLP PIPELINES & ISRO GEOSPATIAL AGENTS
FORECAST
DETERMINISTIC EXTRACTION · ZERO PII LEAKAGE · <85MS INFERENCE LATENCY
3:4:2 NET
“We can only see a short distance ahead, but we can see plenty there that needs to be done.”
Alan Turing — Computing Machinery and Intelligence, 1950
STACK TICKER
SURYA OCR ▲ PROD CHANDRA OCR ▲ REGIONAL BYT5 NLP ▲ FINE-TUNED YOLOV8 ▲ REAL-TIME FASTAPI ▲ ASYNC LANGCHAIN ▲ AGENTIC OPENCV ▲ VISION PYTORCH 2.4 ▲ CORE GEOPANDAS / GDAL ▲ GEOAI PII MASKING ▲ SECURE POSTGRESQL ▬ DB SURYA OCR ▲ PROD CHANDRA OCR ▲ REGIONAL BYT5 NLP ▲ FINE-TUNED YOLOV8 ▲ REAL-TIME FASTAPI ▲ ASYNC LANGCHAIN ▲ AGENTIC OPENCV ▲ VISION PYTORCH 2.4 ▲ CORE GEOPANDAS / GDAL ▲ GEOAI PII MASKING ▲ SECURE POSTGRESQL ▬ DB
01

Dossier

100%
PII MASKING & STRUCTURED EXTRACTION ACCURACY
99%+
REGIONAL OCR TEXT NORMALIZATION WITH BYT5
ISRO’25
CHAIN-OF-THOUGHT AGENTIC GEOSPATIAL REASONING
7.6/10
B.E. COMPUTER SCIENCE & ENGINEERING (LJIET)

AI/ML Engineer specializing in Intelligent Document Processing (IDP), OCR pipelines, Computer Vision, and Agentic Generative AI systems. At Proplegit, I build end-to-end document intelligence architectures for complex Maharashtra Government records and regional Gujarati newspapers—replacing cost-prohibitive cloud APIs with optimized in-house Surya and Chandra OCR pipelines, ByT5 sequence correction, and automated PII redaction.

Beyond document AI, I engineer multi-step reasoning systems and spatial intelligence workflows: from building the Chain-of-Thought Geospatial Reasoning System for the ISRO BAH 2025 competition with LangChain, GeoPandas, and GDAL, to real-time object detection and decision engines in QueenKitchen using YOLOv8 and FastAPI. Every system is built with production rigor—modular architectures, automated validation schemas, deterministic error recovery, and robust REST APIs.

02

Services

Structured engineering engagements tailored for applied AI systems, model deployment, and computer vision pipelines.

INTERACTIVE BRIEF BUILDER

Configure your project requirements below. The interactive panel generates a clean, structured technical brief ready to send or copy.

STEP 01 · SELECT FOCUS DOMAIN
STEP 02 · ENGAGEMENT MODEL
STEP 03 · ESTIMATED TIMELINE
STEP 04 · CONTEXT & DETAILS
TERMINAL BRIEF OUTPUT REF/20261012
SEND VIA EMAIL →
SCHEDULE TECHNICAL INTRO CALL → EXPLORE OPEN SOURCE REPOSITORIES ↗
03

Field Notes

Direct engineering post-mortems from production AI and systems: real constraints, deep diagnostics, interventions, and measured outcomes.

14×
CASE STUDY · AUGUST 2026

The Spatial Projection Mismatch That Strangled Batch Inference

PROBLEM

In the ISRO Bhuvan satellite spatial risk pipeline (ISRO BAH 2025), high-resolution parcel polygons were ingested to query multimodal satellite indices. During benchmark load testing, batch inference throughput collapsed by an order of magnitude: a 2,000-parcel batch took over 180 seconds to compute zonal statistics, causing API timeouts and worker starvation.

APPROACH

Profiling CPU flamegraphs revealed that over 88% of execution time was spent in on-the-fly CRS re-projections (EPSG:4326 to UTM EPSG:32643) inside a tight Python loop for every single raster window slice. Rather than converting geometries per-tile, we restructured the ingestion pipeline: parcels are bulk-transformed into native raster projection space once, spatial indices are built in PostGIS using geometry bounding-box centroids, and window reads are chunked using GDAL block cache boundaries.

RESULT

Batch calculation time dropped from 184 seconds down to 13.1 seconds—a 14× throughput acceleration with zero loss in geometric precision. Peak memory utilization decreased by 62% because intermediate GeoDataFrames are no longer replicated across multi-process workers.

LESSON LEARNED

In geospatial and deep learning systems, spatial coordinate alignment is often treated as an incidental data cleaning step. In high-concurrency production, implicit on-the-fly reprojection is a silent bottleneck. Pre-aligned coordinate arrays and spatial indexing belong in the data layer, not in inference workers.

99%
CASE STUDY · JULY 2026

Eliminating Hallucination Drift in Structured Agentic Extraction

PROBLEM

In an automated document intelligence pipeline processing unstructured municipal survey forms, an LLM agent was tasked with extracting tables and legal easements into strict JSON schemas. Prompting the model with few-shot examples yielded a 22% failure rate due to schema drift, missing nested fields, and subtle hallucinated dates.

APPROACH

Replaced open-ended completions with strict grammar-guided decoding (constrained sampling via JSON schema BNF grammars) coupled with an OCR layout-aware token bounding box prompt. Added an automated dual-pass validation loop: a fast local validator verifies regex expressions and mathematical checksums before committing to the database.

RESULT

Schema conformance reached 99.4% on automated evaluation across 4,500 real-world scan samples. The need for manual human escalation was confined exclusively to degraded low-DPI scans, reducing processing turnaround from 48 hours to 18 seconds.

LESSON LEARNED

Never rely on prompt enthusiasm to enforce schema guarantees. Deterministic grammar constraints at inference time guarantee structural validity, allowing the model's reasoning capacity to focus solely on entity resolution.

CORRECTIONS DESK

Newspapers publish corrections. Senior engineers publish post-mortems.
An official record of flawed early assumptions, detected failure modes, and permanent architectural rules.

CORRECTION · 01
INITIAL ASSUMPTION:

Claimed a 0.96 validation IoU for remote sensing satellite parcel segmentation based on uniform random spatial pixel train/validation partitioning.

PRODUCTION REALITY:

Random uniform splitting leaked spatial autocorrelation between adjacent GeoTIFF tiles. Real generalization dropped to 0.74 on unseen satellite swaths until coordinate-buffered spatial block cross-validation was enforced.

↳ ARCHITECTURAL RULE: Uniform random splitting on spatial or temporal data guarantees self-deception. Always enforce coordinate-buffered block partitions.
CORRECTION · 02
INITIAL ASSUMPTION:

Benchmarked isolated single-stream TensorRT vision inference at 14ms and assumed 50 concurrent requests would scale linearly to 3,500 req/sec.

PRODUCTION REALITY:

Single GPU context locks and host-to-device PCIe memory contention degraded P99 latency past 180ms under concurrency. Solved by dynamic micro-batching queues and dedicated CUDA execution streams.

↳ ARCHITECTURAL RULE: Single-stream benchmark figures are academic fiction. Measure saturation throughput under concurrent load before declaring SLAs.
CORRECTION · 03
INITIAL ASSUMPTION:

Relied on prompt instructions (“stop reasoning once the query is resolved”) to govern multi-agent LangGraph reflection and tool recursion.

PRODUCTION REALITY:

Under ambiguous edge queries, agents looped in recursive self-reflection, burning $42 in API tokens in 8 minutes. Enforced hard state-machine step caps, cycle-detection heuristics, and deterministic schema validators.

↳ ARCHITECTURAL RULE: Prompt-based stopping conditions are polite suggestions; deterministic state-machine circuit breakers are guarantees.
04

Magazine

Editorial essays and technical analyses on AI engineering, inference architectures, and computer vision systems.

N.03
SEPTEMBER 2026 · GEOSPATIAL AI

Mitigating Urban Heat Islands with Satellite Remote Sensing & Multimodal Vision

READ
OVERVIEW

Urban Heat Islands (UHI) disproportionately affect dense metropolitan regions, driving peak energy demands and public health vulnerabilities. By fusing thermal infrared satellite data (Landsat 8/9 TIRS) with high-resolution multispectral imagery (Sentinel-2), our model predicts surface temperature micro-anomalies and correlates them with localized canopy coverage and albedo metrics.

TECHNICAL ARCHITECTURE

We deployed a dual-stream convolutional architecture combined with feature pyramid networks (FPN). The spatial branch captures fine-grained urban geometry (asphalt, concrete, rooftop materials), while the spectral branch models normalized difference vegetation index (NDVI) and normalized difference built-up index (NDBI).

RESULTS & POLICY IMPACT

The model achieved an R² score of 0.89 in predicting micro-climate surface temperatures across dense urban grids. Municipal planners can simulate targeted green roofing and cool-pavement interventions before allocating infrastructure capital.

N.02
AUGUST 2026 · AGENTS & REASONING

Beyond Simple Prompt Chains: Building Deterministic State Graphs for Autonomous Agents

READ
OVERVIEW

Linear prompt chaining fails whenever real-world systems encounter partial failures, API timeouts, or ambiguous outputs. Transitioning from unstructured ReAct chains to explicit cyclic state graphs (using LangGraph / state machine topologies) transforms experimental prototypes into resilient enterprise systems.

DETERMINISTIC GUARDRAILS

By restricting transitions to explicit graph edges and enforcing state validation schemas at every node, agentic execution becomes reproducible, debuggable, and auditable. State rollbacks and human-in-the-loop escalation paths are codified directly in the graph definition.

PRODUCTION INSIGHT

Agents should not decide their own operational boundaries. Give them autonomy over reasoning steps, but enforce strict topological constraints on the execution graph.

N.01
JULY 2026 · INFERENCE ENGINEERING

Quantization, Paged Attention & Speculative Decoding: High-Throughput Serving on Constrained Hardware

READ
OVERVIEW

Serving modern 70B+ parameter models on commodity or single-node GPU clusters requires deep optimization across memory bandwidth and arithmetic intensity. Paged attention eliminates KV-cache fragmentation, while activation-aware quantization (AWQ) preserves perplexity within 0.1% of FP16 baselines.

BENCHMARKS

Combining 4-bit weight quantization with draft-model speculative decoding yielded a 3.4× speedup in tokens per second per dollar, turning cost-prohibitive proprietary API dependencies into sustainable on-premise infrastructure.

05

Projects

ONGOING (2025) PYTHON · AGENTIC LLMS · RAG · GEOAI · LANGCHAIN · GEOPANDAS · RASTERIO · GDAL

An autonomous agentic LLM architecture that translates ambiguous natural language requests into multi-step, executable geospatial workflows. Combines chain-of-thought reasoning with RAG over GIS tools including GDAL and QGIS to synthesize and execute automated spatial processing pipelines.

PROBLEM

Remote sensing analysts and GIS practitioners frequently face complex, multi-stage data processing tasks—such as reprojecting satellite rasters, extracting land-use polygons, and computing zonal statistics. Manually chaining GDAL, Rasterio, and QGIS scripts is slow, brittle, and inaccessible to non-expert users.

APPROACH

Architected a Chain-of-Thought agentic framework in Python using LangChain and tool-augmented LLMs. The system decomposes natural language spatial queries into topological task graphs, retrieves relevant API primitives and geospatial coordinate references via RAG, and generates executable GeoPandas/GDAL code pipelines with automated schema validation.

RESULT

Autonomous generation of end-to-end geospatial workflows with deterministic step execution and error recovery loops. Successfully validated on complex spatial operations across satellite imagery and vector datasets for the ISRO BAH 2025 challenge.

PRODUCTION ENTERPRISE SURYA OCR · NLP · PII MASKING · FASTAPI · PYTHON

High-throughput intelligent document processing pipeline developed at Proplegit for official Maharashtra Government documents. Handles noisy low-DPI scans, dense stamps, and complex multi-column layouts, performing structured data extraction, text normalization, and automated PII masking.

PROBLEM

Official municipal and land records in Maharashtra arrive as historical physical scans filled with heavy ink stamps, background watermarks, skewed tables, and sensitive citizen identities. Previous reliance on cloud LLM APIs (Gemini) suffered from steep operational costs, high latency, and privacy compliance hurdles.

APPROACH

Built an in-house document intelligence pipeline based on Surya OCR. Engineered custom layout-aware text segmentation, deskewing filters, regex-based validation checksums, and strict PII redaction algorithms for Aadhaar, phone, and registration numbers. Fully replaced the external Gemini dependency with an optimized local NLP pipeline.

RESULT

Drastically reduced processing costs and API latency while guaranteeing 100% on-premise PII masking compliance. Delivered robust, validated JSON outputs from degraded government scans directly into Proplegit's enterprise database.

PRODUCTION AI & NLP CHANDRA OCR · BYT5 · INDICXLIT · PYTORCH · HUGGINGFACE

Regional-language newspaper digitization pipeline engineered at Proplegit. Leverages Chandra OCR for complex Indic scripts, a fine-tuned ByT5 sequence model for OCR text correction and normalization, and IndicXlit transliteration for automated Gujarati-to-English script conversion.

PROBLEM

Regional Indic scripts, specifically Gujarati newsprint, present unique OCR challenges: conjunct consonants, complex diacritical vowel markers (matras), variable newsprint ink bleed, and lack of standardized open-source OCR tools, resulting in severe character recognition failure rates.

APPROACH

Integrated Chandra OCR specifically tuned for regional Indian typography. To resolve residual OCR distortions, fine-tuned Google's token-free byte-level ByT5 model on paired noisy/clean Gujarati text corpora for character correction and normalization. Built automated transliteration pipelines using IndicXlit for seamless script mapping.

RESULT

Achieved high-accuracy text transcription on dense, multi-column vintage Gujarati newsprint. The searchable digital text archive feeds directly into downstream indexing, entity extraction, and content intelligence engines.

FULL-STACK AI PYTHON · YOLOV8 · OPENCV · FASTAPI · REACT

Computer-vision-based food inventory intelligence platform and decision engine built to minimize food waste. Detects ingredients in real-time using custom YOLOv8 object detection, calculates expiration risk curves, and serves predictive recommendations through an async FastAPI backend and React frontend.

PROBLEM

Commercial and household kitchens lose significant resources to unmonitored ingredient spoilage. Manual inventory logging is tedious and error-prone, leaving kitchens without timely notifications before food perishes.

APPROACH

Trained and optimized a YOLOv8 convolutional detector via OpenCV to perform real-time multi-class ingredient identification from live camera feeds. Developed an algorithmic AI decision engine that predicts freshness lifecycles and suggests dynamic recipe usage. Integrated the model with an asynchronous FastAPI REST API and an intuitive React interface.

RESULT

Fast, real-time bounding box detection with sub-100ms inference on standard hardware. Fully connected end-to-end system that tracks dynamic food inventory and automated waste prevention alerts.

06

Career

JAN 2026 — PRESENT PROPLEGIT · AHMEDABAD, INDIA

AI Engineer Intern

Engineering production-grade Intelligent Document Processing (IDP), regional OCR, and NLP systems.

Maharashtra Government Document OCR Pipeline

  • Built Surya OCR-based pipelines for Maharashtra Government documents, handling noisy scans, watermarks, and complex layouts.
  • Developed OCR post-processing, structured extraction, text normalization, validation, and automated PII masking workflows.
  • Replaced Gemini-based cloud extraction with optimized in-house OCR and NLP pipelines, eliminating recurring API costs and latency.

Gujarati Newspaper Digitization and NLP Pipeline

  • Developed Gujarati newspaper digitization using Chandra OCR for regional-language documents.
  • Fine-tuned ByT5 for text correction, normalization, and OCR post-processing.
  • Built IndicXlit transliteration workflows for Gujarati-to-English script conversion.
Surya OCR Chandra OCR ByT5 IndicXlit PII Masking FastAPI PyTorch HuggingFace
07

Education

SEPT 2023 — AUG 2027 AHMEDABAD, GUJARAT, INDIA
ACADEMIC STANDING 7.6/10 CGPA

Bachelor of Engineering in Computer Science and Engineering

LJ INSTITUTE OF ENGINEERING AND TECHNOLOGY

Rigorous academic foundation in core computer science, systems engineering, and advanced mathematical foundations. Active solver and analytical practitioner of graduate-level Machine Learning examination papers (University of Pennsylvania CIS 520 & GATE Data Science and AI curriculum) spanning convex optimization, Lagrangian duality, SVD/PCA spectral analysis, probabilistic graphical models, and reinforcement learning.

CORE COMPUTER SCIENCE & SYSTEMS
  • Data Structures & Algorithms in Java (Arrays, Linked Lists, Trees, Graphs, Hash Maps)
  • Object-Oriented Programming (OOP) & Software Design Patterns
  • Database Management Systems (Relational Design, SQL, PostgreSQL, MySQL)
  • Operating Systems, Memory Management & Linux Kernel Concepts
APPLIED AI & PRODUCTION PIPELINES
  • Intelligent Document Processing (IDP), Surya & Chandra OCR Pipelines
  • Transformer NLP (ByT5 Fine-Tuning & IndicXlit Script Normalization)
  • Computer Vision (YOLOv8 Real-Time Object Detection & OpenCV)
  • Asynchronous Architectures (FastAPI REST APIs & Agentic LLM Orchestration)
THEORETICAL ML & PAPER SOLVING
  • Convex Optimization, Lagrangian Duality, KKT Conditions & Mercer Kernels
  • SVD & Spectral PCA (Eigenvector Variances, Linear Autoencoder Duality)
  • Probabilistic Graphical Models (Bayesian Networks, d-separation, Markov Blankets)
  • Statistical Learning Theory (Bias-Variance, MLE vs. MAP, MDPs & Q-Learning)
ACADEMIC RIGOR · GRADUATE ML PROBLEM SOLVING

Advanced Machine Learning Theory & Exam Papers Analysis

Active solver and analyzer of graduate-level Machine Learning problem sets, including University of Pennsylvania CIS 520 (Machine Learning) Finals & Midterms (2017–2022) and the GATE Data Science & Artificial Intelligence (DSAI) curriculum.

OPTIMIZATION & DUALITY

Convex optimization, Lagrangian duality, Karush-Kuhn-Tucker (KKT) conditions, Mercer kernel proofs for SVMs, and L1 (Lasso) vs. L2 (Ridge) regularization dynamics.

SPECTRAL & MATRIX THEORY

Singular Value Decomposition (SVD), left/right singular vector duality with AAᵀ and AᵀA, PCA reconstruction error minimization, and covariance eigenvector analysis.

PROBABILISTIC GRAPHICAL MODELS

Bayesian network factorizations, d-separation algorithms, conditional independence queries (C ⊥ D | F), Markov blankets, and minimal parameter bounds for joint distributions.

STATISTICAL LEARNING & RL

Bias-variance decomposition under varying hypothesis complexity, Maximum Likelihood (MLE) vs. MAP estimation, Markov Decision Processes (MDPs), Bellman optimality, and Q-learning.

B.E. Computer Science CGPA: 7.6 / 10 UPenn CIS 520 ML Papers GATE Data Science & AI Convex Optimization Lagrangian Duality (SVM) SVD & Spectral PCA Bayesian Networks (d-sep) Markov Decision Processes Data Structures & Algorithms Java OOP Python Linux
08

Stack

OCR & Document AI

Surya OCR, Chandra OCR, ByT5 sequence modeling, IndicXlit transliteration, OCR post-processing, Intelligent Document Processing (IDP), PII masking, layout parsing, and text normalization.

Generative AI & Agents

Agentic LLMs, Chain-of-Thought reasoning, LangChain, RAG architectures, Prompt Engineering, Structured JSON grammars, and deterministic evaluation benchmarks.

Computer Vision

OpenCV, YOLOv8, Image Processing, Object Detection, bounding box extraction, morphological transformations, contour detection, and camera stream ingestion.

AI & Deep Learning

PyTorch, HuggingFace Transformers, ByT5 fine-tuning, Scikit-learn, neural network architectures, and sequence-to-sequence modeling.

Data Science & ML

NumPy, Pandas, Scikit-learn, Data Preprocessing, Feature Engineering, Exploratory Data Analysis (EDA), statistical evaluation, and model validation.

Backend & APIs

FastAPI, REST APIs, JSON Schema validation, Asynchronous ASGI microservices, API Integration, and modern web development with React and JavaScript.

Geospatial & GIS Tools

GeoPandas, Rasterio, GDAL, QGIS, GeoAI workflows, remote sensing satellite rasters, and spatial vector geometries.

Databases & Systems

PostgreSQL, MongoDB, MySQL, Git version control, Linux environments, Docker containerization, and automated software testing.

ML Theory & Foundations

Convex optimization, Lagrangian duality & KKT conditions, SVD/PCA spectral analysis, Mercer kernels, Probabilistic Graphical Models (d-separation), MLE/MAP inference, and Markov Decision Processes (MDPs / Bellman optimality). Solved UPenn CIS 520 & GATE DSAI papers.

Continuous Engineering Verification & Verified Professional Certifications

VERIFIED CREDENTIAL IBM · COURSERA Introduction to HTML, CSS, and JavaScript Skills: HTML5, CSS3, Modern JavaScript ES6, DOM Manipulation, Web Architecture VERIFY CREDENTIAL ↗ VERIFIED CREDENTIAL UNIVERSITY OF PENNSYLVANIA · COURSERA Inheritance and Data Structures in Java Skills: Object-Oriented Programming, Inheritance, Polymorphism, Linked Lists, Trees, Graphs VERIFY CREDENTIAL ↗ VERIFIED CREDENTIAL IBM · COURSERA Exploratory Data Analysis for Machine Learning Skills: Data Cleaning, Feature Engineering, NumPy, Pandas, Statistical Analysis, Hypothesis Testing VERIFY CREDENTIAL ↗ VERIFIED CREDENTIAL DEEPLEARNING.AI / COURSERA Building Generative AI-Powered Applications with Python Skills: LLMs, Prompt Engineering, RAG Systems, LangChain, Python AI Frameworks, Model Evaluation VERIFY CREDENTIAL ↗
09

Security

This portfolio audits itself. The table below inspects production security headers, CSP guarantees, and AI output handling policies.

$ security headers & safeguards

SYSTEM AUDIT · LIVE
Content-Security-Policy ON default-src 'self'; script-src 'self' 'unsafe-inline' https://fonts.googleapis.com; style-src 'self' 'unsafe-inline' https://fonts.googleapis.com; font-src 'self' https://fonts.gstatic.com;
Strict-Transport-Security ON max-age=31536000; includeSubDomains; preload
X-Content-Type-Options ON nosniff
Referrer-Policy ON strict-origin-when-cross-origin
OWASP LLM02 · Output Handling ON Sanitized HTML output in terminal; model responses treated as untrusted text
CSP: no 'unsafe-eval' ON Strict execution policy enforced

Terminal interactions sanitize inputs and model outputs. Refusing inline evaluation protects client sessions against prompt injection escalation.

10

Booking

Schedule a 30-minute introductory discussion. We'll explore your machine learning challenges, architecture requirements, or project roadmap with zero obligation.

TECHNICAL INTRO CALL · 30 MINUTES

Directly connect with Krish Patel to discuss AI/ML development, RAG systems, or computer vision implementations.

REQUEST A MEETING VIA EMAIL →