Pallav Nag | Backend & AI Engineer

AI_LAB · ACTIVE

What I'm Building

Flagship · Quant + AI

Systematic Equity
Research Engine

NSE quant stack · survivorship-free, point-in-time backtested

The problem

Most backtests flatter whoever built them. I wanted an Indian-equity research stack where an idea only ships if it beats a random portfolio drawn from the same pool, on the same dates, through the production code path.

The approach

Every trading evening it ingests NSE data into SQLite, rebuilds a point-in-time universe, ranks it with a proprietary cross-sectional signal and emails the day's orders. A batched Gemini pass tags news as a downside-only veto. The research harness shares production's code path, so the backtest can't drift from what actually trades.

Architecture

NSE bhavcopy + corporate actions → SQLite · 3.5M rows

↓

split / holiday / delisting / fund cleaning → point-in-time universe

↓

proprietary ranker ◆ Gemini news veto ◆ risk caps

↓

orders email → positions ledger, graded vs random + index

How it works

Ingest: 7 years of NSE data, 3,695 symbols including delisted ones, written through idempotent upserts.
Clean: survivorship, raw splits, phantom holiday sessions, look-ahead: each bias measured, then fixed. A 14-defect audit before publishing.
Rank: one proprietary signal; news can veto a pick but never add one.
Prove it: every idea vs a 5-draw random control and a path-noise null; CAGR averaged over 675 simulated paths. Roughly 1 in 70 ideas survive.
Run: unattended via launchd, with a positions ledger and broker-side stop orders.

Results · backtested, net of cost + tax

+4.77pp

Edge / pick vs random

+13.0

t-stat, per pick

24.6%

CAGR after tax

CAGR after tax, Jun 2022 to Oct 2026

Ranked book

24.6%

10th pct path

15.8%

Nifty 500

10.5%

Random picks

~10.5%

Survivorship-free point-in-time NSE panel, production code path, next-day-open fills, 0.5% round trip, STCG/LTCG + cess netted per financial year. CAGR is the mean of 675 simulated paths; median max drawdown −30%. Caveats, stated plainly: one bull-market window, zero live trades.

PythonpandasNumPySQLiteNSE data StatisticsGeminiBacktestinglaunchd

Private tool · request a walkthrough → · read the write-up →

AI Agents · Distributed

GraphReview:
Multi-Agent PR Review

Kafka · Kubernetes · GraphRAG · Claude

The problem

A diff-only reviewer can't see the caller two files away that a change just broke. Feeding it whole files instead burns tokens on code that doesn't matter and still misses what does.

The approach

Specialist Claude agents (correctness, security, tests) plus a linter review every PR in parallel off one Kafka topic. Instead of reading whole files, they investigate through tools backed by a Graphify code graph and a hybrid Qdrant index. An aggregator merges duplicates, an LLM-as-judge critic drops false positives, and one GitHub review lands with inline comments.

Architecture

GitHub PR→gatewayHMAC · dedupe

pr.review.requested · fan-out, 4 consumer groups

staticruff reviewercorrectness securityappsec testscoverage

⇄ tool calls · code graph (Redis) + dense + BM25 (Qdrant)

review.agent.results · keyed by job_id

aggregator→critic→publisher

one GitHub review · inline comments

Kafka topiccalls Clauderetrieval

How it works

Gateway: verifies the webhook HMAC, dedupes redeliveries per head SHA, publishes one job.
Agents: one consumer group per role; a bounded tool loop starts from a graph impact report (changed symbols → callers → tests).
GraphRAG tools: callers and tests from the graph; similar code and team conventions from hybrid search fused with RRF.
Aggregator → Critic: merges duplicates, boosts confidence when agents agree; the judge drops false positives.
Publisher: one review; broken callers outside the diff land in the body, not lost.

Results · 40-PR burst on Kubernetes

1→6

Pods in ~20 s (KEDA)

~50s

Backlog drained

4

Agents in parallel

Single-node kind cluster (6 vCPU / 10 GB), offline LLM at 2 s simulated latency, KEDA scaling each agent role on its own consumer lag. Review quality is scored on an 11-case seeded-bug benchmark: graph_rag vs full_files vs diff_only.

PythonKafkaKubernetesKEDAStrimzi Claude APIGraphRAGGraphifyQdrantRedis FastAPIMCPPrometheus

Read the architecture deep-dive View on GitHub

Measure the edge, don't trust it

A survivorship-free backtest and a random-portfolio control beat conviction. If it can't beat a dartboard, it doesn't ship.

Less context, more signal

Graph retrieval and hard data-quality filters over brute force. Smaller, sharper inputs beat feeding everything in.

Specialised > generalised

One job per agent, one responsibility per module. Coordination beats monoliths, in AI pipelines and trading systems alike.

WHERE_I_SHIP · PRODUCTION

Day Job

Software Developer 1

Amadeus IT Group · Bengaluru

Jul 2024 - Present

★ Spot Award Q3 2025 · team of 40 ↑ €172K saved / year ↓ 70% security vulns ↓ 70–80× AI tokens

Own backend systems for airline departure control and flight management at global scale. Independently traced and fixed a critical SLA breach now saving €150K/yr, cut a further €22K/yr from log-processing overhead, and automated a manual reporting workflow that returned 90 developer-days to the team. Drove org-wide adoption of AST-based context tooling (Graphify) for a 70–80× drop in AI token usage, and added BlackDuck + Fortify gates to CI/CD that cut security vulnerabilities by 70%. Primary POC for a zero-impact quarterly data migration across teams.

JavaC++KafkaMongoDB EdifactSpring MVC/BootC++ Templates REST APIsPodmanAWSCI/CD

Software Developer Intern

Amadeus IT Group · Bengaluru

Feb 2024 - Jul 2024

Designed and built an event-driven data pipeline streaming real-time passenger & baggage data from check-in to flight management, replacing polling with live CDC. Architected on Kafka Connect + Debezium + MongoDB; containerised with Podman. Presented at internal showcase; outcomes shipped to production roadmap.

Java / JAX-RSKafka ConnectDebezium CDC MongoDBPodmanMaven

SIDE_WORK · OPEN_SOURCE