DATA → AGENTS → PRODUCT
AI Engineer at Facets Gems Polishing (K. Girdharlal), building systems that chain AI calls together, own the databases underneath them, and ship with security layers baked in — not bolted on. Diamond-industry data by trade, agentic engineering by practice, fluent across Claude, Gemini/Antigravity and the rest of the modern AI stack.
CLARITY
HOW HE THINKS
Trained on pure mathematics — probability, statistics, linear algebra, number theory — before it ever touched a codebase. Spent a year designing adversarial prompts to find exactly where AI models reason incorrectly, then applying the same scrutiny to his own pipelines. Clarity here means: distrust the output until the method is checked.
- 500+
- advanced math problems solved, Chegg
- 200+
- model outputs adversarially evaluated
CUT
HOW HE BUILDS
As AI Engineer at Facets Gems Polishing (K. Girdharlal), the job is rarely a single prompt — it's graphs: custom-orchestrated, multi-agent systems where one step hands its output to the next, a role-based agent (CrewAI-style) picks up specialized work, and the path branches conditionally instead of running straight through. Built on FastAPI / REST services underneath, backed by both SQL and NoSQL databases he owns end to end — schema, integrity, performance — and wrapped in security layers he deploys himself: authentication and access control, encryption, and API/network hardening. Outside work, the same instinct shows up in the Financial Tracker's agent that reads an SMS, parses it, categorizes it, and acts — mostly on-device. The cut is precise: automate the repeatable, keep humans on the judgment calls.
- 60–70%
- manual effort removed via automated pipelines
- monthly → weekly
- reporting cadence shift
CARAT
SCALE OF THE WORK
Diamond lifecycle data doesn't forgive sloppy validation — a single bad record can misprice a stone. Carrying that discipline into every dataset since: high record counts, high accuracy bars, no shortcuts on data quality.
- 300K+
- records validated across the diamond lifecycle
- 100K+
- carat stones tracked
- ~99%
- data accuracy achieved
COLOR
BREADTH ACROSS PLATFORMS
Comfortable moving between classical data tooling and frontier AI tooling in the same afternoon: SQL and NoSQL databases, REST/FastAPI services and Power BI on one side, multi-agent orchestration and front-end AI deployment on the other. Builds and ships across Claude, Gemini/Antigravity, and the surrounding agentic toolchain — not a spectator to the AI-tooling shift, an active user of it.
- Python · FastAPI · SQL / NoSQL
- core engineering stack
- Claude · Gemini/Antigravity
- AI development platforms
- Auth · Encryption · API hardening
- security layers deployed
REPORT
CERTIFIED RECORD OF EMPLOYMENT
Apr 2025 — Present
AI Engineer — Facets Gems Polishing Pvt. Ltd. (K. Girdharlal)
- Automates core business workflows with AI across the diamond lifecycle — inventory, pricing and reporting
- Designs multi-agent, custom-orchestrated systems (CrewAI-style role-based agents, conditional hand-offs) rather than single-shot prompts
- Builds the REST / FastAPI services those systems run on
- Owns the production databases behind them, both SQL and NoSQL — schema, integrity, performance
- Deploys security layers protecting data pipelines and internal tooling: authentication and access control, encryption, and API/network hardening
- Manages inventory data across 100K+ carat stones, 300K+ records, validated to ~99% accuracy
- Automated Python pipelines, cutting manual effort 60–70% and moving reporting from monthly to weekly
- Produces P&L reports that speed up pricing and inventory decisions
- Integrated live data into Flask dashboards for real-time reporting
Oct 2024 — Mar 2025
AI Model Reviewer — Outlier
- Designed adversarial prompts to evaluate and improve AI model performance
- Applied statistical reasoning across 200+ model-output evaluation tasks
- Specialized in mathematics, probability and statistics validation
- Reviewed 50+ responses for accuracy and logical consistency
Jun 2023 — Nov 2024
Subject Matter Expert — Chegg Inc.
- Solved 500+ advanced mathematics problems spanning probability and statistics
- Delivered structured, data-driven solutions under a strict correctness bar
- Maintained a 4.2+ rating across 100+ pieces of positive feedback
APPENDIX / EDUCATION
Nov 2022 — Jun 2024
M.Sc. Computer Science
Banaras Hindu University
Jul 2019 — May 2022
B.Sc. Mathematics (Honors)
Maharaja Sayajirao University of Baroda
APPENDIX / CERTIFICATIONS
- Google Data Analytics Professional CertificateCoursera
- Data Science Specialization (R)Coursera
- Natural Language ProcessingNPTEL
- AI-Data Science ProgrammeNASSCOM
SPECIMENS
SIX SYSTEMS, EXAMINED UNDER LOUPE — ONE STILL SEALED
VittaraFinOS
Flutter · On-device ML · Encrypted-first
“Track Wealth, Master Life” — a local-first personal finance app for Android. No backend, no server, no cloud plaintext: every transaction, balance and insight lives in an encrypted database on the device itself.
- ·SQLCipher-encrypted database; PIN (PBKDF2-HMAC-SHA256, 120k iterations) + biometric app lock
- ·On-device ML captures transactions automatically: SMS bank-alert parsing, statement OCR, voice commands
- ·AES-256-GCM encrypted export/backup and restore, device-to-device sync
- ·CI-enforced security hygiene — secret/backup scans, no raw logging, no silent empty catches
Freight Transport Management System
Next.js · Express · React Native
A full-stack TMS running a real freight business end to end — consignment booking through dispatch, invoicing, payments and ledgers. Built and shipped for a live logistics client; described here as a technical case study, client name withheld.
- ·Enforced consignment lifecycle — Booked → Loaded → In Transit → Delivered → Settled — with e-way bills, challans and POD uploads
- ·GST invoices generated as PDFs (Puppeteer + HTML templates), party-wise billing and payment tracking
- ·JWT auth in httpOnly cookies, role-based access control, an agent portal, and public no-login shipment tracking
- ·React Native (Expo) companion app; whole stack ships as one deployment
Invoicing & Receivables System
Supabase · Business Ops
A multi-page business-operations tool for a real logistics client — invoice entry with A4 print/PDF export, party-master maintenance, and a dedicated payments and receivables register. Client name withheld.
- ·Invoice entry, print/PDF export, and bulk invoice import via a Supabase Edge Function
- ·Party-master maintenance and lookup, plus a dedicated receivables register
- ·Command-center dashboard with recent totals and quick navigation
Legal Text Summarizer
NLP · BERT · T5
Extractive and abstractive summarization over real legal documents — trained and compared BERT and T5 approaches on a legal corpus, with a full preprocessing pipeline behind them.
- ·BERT for extractive summarization, T5 for abstractive
- ·Tokenization, lemmatization, stopword removal pipeline
- ·Head-to-head performance comparison across models
Diamond Lifecycle Analytics Pipeline
Python · SQL · Flask
The production system behind the CARAT numbers above: ingestion, validation and cleaning for six-figure record volumes, feeding live P&L and inventory dashboards used for real pricing decisions.
- ·300K+ records validated to ~99% accuracy
- ·Pipelines cut manual reporting effort 60–70%
- ·Real-time Flask dashboards for pricing and inventory
Redacted
Real-world deployment — approach shown, client and implementation withheld
- CONTEXT
- Sports analytics — ball-by-ball match data across multiple competitive formats.
- PROBLEM
- Raw match-event data doesn't answer questions by itself. Turning it into analysis that's actually decision-useful means slicing it across many dimensions at once — time, players, teams, venues, phase of play — without the tool collapsing under its own complexity.
- APPROACH
- A multi-format analytics workspace: dedicated processing pipelines per format feeding a shared statistics layer, with dashboards built for match-center detail down to season-level and career-level views.
READOUT
NOT SCREENSHOTS — LIVE INSTRUMENTS. MATH, CS, AND A REASON TO STAY.
Fetching real commit activity…
Each tick draws n independent uniform[0,1] values and plots their mean. Drag n up and watch an unremarkable flat distribution collapse into a bell curve — the theorem his stats work leans on daily, running live in your browser.
- observed mean
- 0.0000
- observed sd
- 0.0000
- theoretical sd (1/√(12n))
- 0.1021
Drag the matrix. Watch the plane, the unit square, and the eigenvectors move live — the same linear algebra behind his “linear algebra” line, running in your browser instead of sitting on a resume.
- determinant
- 1.000
- trace
- 2.000
- eigenvalues
- λ₁ = 1.00, λ₂ = 1.00
- reading
- orientation preserved, area ×1.00
The unit circle, unrolled: e^(iθ) = cos θ + i sin θ. Trigonometry and the exponential function are the same object, viewed from two angles.
- θ
- 0.00 rad (0°)
- cos θ
- 1.000
- sin θ
- 0.000
- e^(iθ)
- 1.00 + 0.00i
watch for θ = π
Where math and computer science actually meet: six growth rates, same n, log-scaled y-axis so they all fit in one frame. Drag n and watch how fast the gap between them explodes.
Click only the prime numbers before the clock runs out. Number theory, timed — not on a resume, running live.
- score
- 0
- streak
- 0
- best streak
- 0
SEAL
THIS REPORT IS CERTIFIED CORRECT AS OF TODAY
Open to roles and collaborations at the intersection of data engineering and agentic AI. Reach out directly, or send a note below.