David Jeremie Anand

I build AI systems, then prove they work.

Applied AI / Agentic Systems / Data and Workflow Solutions

New Delhi, India

View the work jerryanand10@gmail.com Résumé

Scroll

The short version

I make vague systems measurable.

Two years at Genpact turning tangled operations into numbers you could act on — then bringing the same discipline to AI, where almost nobody measures anything. I build agentic RAG, multi-agent reasoning and LLM evaluation, each backed by tests and CI, and I’m currently pursuing a PGP in Applied AI and Agentic Systems at Masters’ Union.

0 Faster turnaround
0 Fewer defects
0 Productivity uplift
0 Less reporting effort

Selected work

Things I built, and why they exist.

01Python · FastAPI · Ollama/Groq · pytest

FirstLine — An AI First Line of Support

I worked in an ops queue for two years. FirstLine is that job rebuilt as an agent: it triages a live incident queue — root causes, SLA clocks, escalation judgment, RCA drafts — on a simulated ops floor that knows the truth behind every ticket it generates.

Which means the agent can be graded, not vibed about. Every shift ends in a scorecard — triage accuracy, missed P0s, escalation precision — because the simulator holds hidden ground truth the agent never sees.

FirstLine · eval real capture
$ firstline run --scenario fog --seed 7 35 tickets · severities lie · shift 480m rules triage 1.000 · esc-prec 1.000 · 100.0 llama3.2 triage 0.857 · esc-prec 0.529 · 81.4 the 3B cried wolf — 17 escalations, 8 needless 0 missed P0s on either backend
768-test CI Live shift dashboard Ollama · Groq · offline rules
Case study →GitHub ↗Live demo ↗
02Python · FastAPI · Claude API · pytest

Day 46 — A Late-Payment Case Builder

A client pays late, and the whole deal lives in a WhatsApp chat. Day 46 reads the chat and the invoice, builds a timeline where every date opens the message behind it, and fills a payment reminder, a demand letter and a case file for an advocate, which the freelancer checks, signs and sends themselves.

The AI only points: it marks the messages that record an order, a delivery or a promise to pay, and an entry survives only if its quote is word for word in the message it cites. It supplies no dates, no amounts and no wording. Code reads those from the messages, and the interest comes from an engine that matches a blind rewrite to the paisa on 2,500 cases.

day46 · accuracy real run · 24 Sep
GET /accuracy · 12 unseen chats 262 events, answers fixed first opus-5 found 99% · right 94% sonnet-5 found 97% · right 95% haiku-4.5 found 78% · right 95% sonnet-5 · one-line prompt found 92% · right 85% 0 of 1,043 quotes hallucinated
1,707-test CI Blind rewrite, to the paisa Masked before the AI sees it
Case study →Live demo ↗
03Python · asyncio · pytest · GitHub Actions

PromptDrift — Prompt Regression Testing

You tweak a prompt to fix one bad case, it works, you ship it — and three other cases broke silently. PromptDrift snapshots prompt behaviour into a committed baseline and fails CI when a change makes things worse.

It refuses to answer “is this prompt good?”, which is unanswerable, and answers “is it worse than yesterday?” — the question that actually blocks a merge. Cheap deterministic checks run first and short-circuit the paid LLM-as-judge calls.

promptdrift · run real capture
$ promptdrift run suite.yaml offline-demo · mock · 4 cases ✓ 3 passed ✗ billing 100% → 0% REGRESSED — expected 'BILLING', got 'OTHER' FAIL 3/4 · exit 1 · merge blocked
157-test CI Python 3.10–3.13 Runs fully offline
View repository ↗
04FastAPI · FAISS

Synapse — Agentic RAG Assistant

BM25 and FAISS fused through Reciprocal Rank Fusion with LLM listwise reranking, feeding a four-stage agent loop of planner, retriever, synthesizer and critic.

The critic is the interesting part: it grounds every answer in retrieved evidence and flags its own unsupported claims.

synapse · critic real capture · llama3.2
query what changed in Q3 policy? draft retention now 365 days [1] applies to all regions [1] critic ✗ not supported — [1] says 180, not 365 ✗ "all regions" — no excerpt → sent back to draft [1] policy_v3.pdf [2] notes.md
31-test CI Render
Case study →GitHub ↗Live demo ↗
05Flask · SQLite · Docker

Ledgr — AI Finance Manager

A full-stack finance tracker with dashboards, budgets and CSV import/export, plus AI categorisation using concurrent LLM calls with a keyword fallback that works with no network at all.

Its assistant calls tools to query real transaction data rather than trusting an LLM to do arithmetic.

ledgr · tool call real capture
user dining spend last month? tool get_totals( start="2026-07-01", end="2026-07-31", category="Food & Dining") → expenses 79.32 · 3 rows reply 79.32 on dining in July the database did the maths
58-test CI Tool-calling
Case study →GitHub ↗Live demo ↗
06SDR · RF and wireless systems

Wireless Temperature Detection — HackRF One

An RF prototype that pulls remotely transmitted temperature signals out of the air and decodes them — the full software-defined radio chain of capture, filtering, demodulation and decoding, validated against a reference sensor.

The waterfall behind the title of this page is a spectrogram, the display this project lives in. It seemed a fitting thing to open with.

Signal analysis Anomaly detection

Experience

Two years inside the machine.

2024Aug 2024 — Aug 2026

Genpact

Associate and Subject Matter Expert — Operations · Gurugram & Noida
  • Analysed high-volume workflows and drove process redesign, contributing to an 18% reduction in turnaround time.
  • Ran structured root-cause analysis of recurring defects and SLA deviations, contributing to a 30% reduction in operational defects.
  • Automated management reporting in Excel and Power BI, cutting recurring effort by 25%.
  • Partnered with Risk, Compliance and Quality on controls and workflow improvements, contributing to a 22% productivity uplift.
  • Documented process gaps, tracked issues through Buganizer and validated corrective actions.

Internships

Where I learned the fundamentals.

2023Jul — Aug 2023

Bhagwan Parshuram Institute of Technology, GGSIPU

Data Analytics Intern · New Delhi

Analysed institutional datasets in Python and Excel, building KPI dashboards and automated reports that cut repetitive manual effort by 25%.

2022Jul — Sep 2022

Defence Research and Development Organisation (DRDO)

Research and Development Intern · India

Supported development, validation and signal analysis of an embedded wireless sensing prototype, improving reliability and reducing errors.

2020Sep — Oct 2020

AGIXURY

Web Development Intern · Remote

Built responsive interfaces from business requirements and improved usability across desktop and mobile.

Education

Electronics, then AI.

2026Present

Masters’ Union Current

PGP in Applied AI and Agentic Systems · Gurugram
2020To 2024

B.Tech, Electronics and Communication Engineering

Bhagwan Parshuram Institute of Technology, GGSIPU · CGPA 8.53 / 10
2018To 2020

Faith Academy Senior Secondary School, CBSE

Class XII 92.8% · Class X 94.2% · New Delhi

Certified

Toolkit

What I reach for.

Programming
PythonSQLFastAPIFlaskData cleaningExploratory analysis
Applied AI
Agentic AIRAGLLM evaluationLarge language modelsPrompt engineeringGenerative AIMachine learning
Systems & cloud
FAISSLLMOpsGenAIOpsAzure AISnowflakeModel APIs
Tooling
GitGitHub ActionsDockerpytestPower BITableauAdvanced Excel
Delivery
Root-cause analysisProcess improvementWorkflow optimisationStakeholder communication

Get in touch

Building something that needs measuring?

I’m looking for applied AI and forward-deployed engineering roles — the kind where you sit with the actual workflow before deciding what to build.

jerryanand10@gmail.com Résumé LinkedIn ↗ GitHub ↗ Field notes

ps — there’s a terminal behind this page. press `