AI Engineering
LoRA fine-tuning, RL environments, synthetic data, evaluation harnesses, model-backed APIs, and LLM integration into real products.
Hyderabad / available for work
I build AI systems that ship to production.
Hi, I am Harshith Pali — an AI engineer and full-stack developer. I build model-backed applications, backend APIs, and web products end-to-end, with a strong bias toward shipping real systems rather than prototypes that never leave a notebook.
0production AI and web systems shipped
Top 300globally · AWS 10,000 AI Ideas Challenge
0tools in active working set
HYDbased in Telangana, India
~/harshith/portfolio
LIVE
> whoami --verbose
harshith pali — AI engineer, full-stack developer, solo builder
github: Harshith029 / huggingface: Harshdev09
> ls ./current-focus
safemigrate-lint [on pypi]
sentinel-agent-security
karma-entity-resolution [hackathon win]
nemotron-reasoning-lora
> cat values.txt
ship real systems, not prototypes
own the whole stack end to end
design for constraints, not for demos
>
RoleAI engineer / full-stack
StatusOpen to work
FocusShipping AI products
01 / About
An AI engineer with a bias toward production.
If I can't ship it, I don't understand it. My best learning happens at the edges — where the model meets the API, where the frontend meets the user, where the idea meets a real deadline.
— Harshith, on how he worksI operate across the full stack of a modern AI product — model training and fine-tuning, backend APIs, cloud infrastructure, and frontend delivery. That breadth is deliberate: the hardest problems in applied AI live at the hand-offs between layers, and solving them requires someone who can hold the whole system in their head.
I pick problems that have real constraints — cost caps, safety requirements, latency budgets, distribution friction. Then I ship working versions and iterate against reality.
Backend & Cloud
FastAPI, Django, Node, Docker, AWS (Lambda, DynamoDB, Bedrock, API Gateway). Production APIs designed for cost and reliability.
Full-Stack Delivery
React, Next.js, Tailwind, vanilla JS. Responsive interfaces and admin dashboards that ship with the backend, not after it.
02 / Selected Work
Production systems with real constraints, built end-to-end.
6featured production builds
5AI / ML systems
WINAI for Bharat 2026 · Govt of Karnataka
TOP 300global rank · AWS challenge
migration lint on PRcheck run
pull request
0042_cleanup.sql
→
2 findings
check run: blocked
branch protection
critical
DROP COLUMN
irreversible data loss.
warning
FK · no NOT VALID
holds AccessExclusiveLock.
parsepglast
analyze33 rules
annotateper line
commenton PR
lock-aware
safe rewrite
MIT
findings by severity
critical · warning · style
safemigrate-lint
A GitHub Action, CLI, and pre-commit hook that lints Postgres migration SQL on every PR — catching the operations that actually break production: lock waits, table rewrites, and unsafe constraints. 33 safety rules + 6 opt-in style rules, parsed with a real Postgres parser (pglast / libpg_query).
Why it matters I scanned ~700 real production migrations from Cal.com, Mattermost, Supabase, Hasura, and TimescaleDB — zero ran the textbook "dangerous" ops the popular linters warn loudest on. The real risks live one layer deeper, and safemigrate-lint is built around those. Published on PyPI, MIT-licensed, free.
RoleSolo engineer
Ships asAction · CLI · hook
Risk handledProd-breaking SQL
Cross-statement context is the trick: a single-statement linter screams about FK-to-new-table and other false positives. safemigrate-lint tracks context across the migration, so it only flags the locks and rewrites that would actually stall a deploy — then posts the safe rewrite inline on the PR.
provenance proxyaction blocked
web_fetchretrieved
summarizetainted
send_emailblocked
provenance set
SYSTEM
USER
AGENT
RETRIEVED
union over the derived_from lineage
authorization
DENY
tainted lineage → high-risk tool call
SENTINEL
A provenance-aware security proxy for agentic AI. It sits in front of an agent as an MCP (Model Context Protocol) server, so every tool call physically traverses it — and authorizes each one against the provenance of everything it derived from. An instruction that arrived inside retrieved web content can never drive a high-risk action, even when it slips past content filters.
The core idea Risk lives in the causal path from instruction-origin to action, not in the text of either. Interception is guaranteed by network topology, not code wrapping — and it's agent-agnostic: it secures Foundry, Claude, GPT, or any custom MCP client with no agent-code changes.
Provenance is a set of trust labels (SYSTEM > USER > AGENT > RETRIEVED_CONTENT) unioned over a span's transitive lineage — not a single mutable enum. A cycle-safe graph walk computes taint; the authorization engine parses policy into a typed AST (never eval) and blocks the action. Taint only clears through an explicit, auditable structured extractor.
solver coverageverified CoT
CIPHER100%
NUMERAL100%
GRAVITY99.7%
Nemotron Reasoning LoRA
A from-scratch, fully-offline pipeline for the NVIDIA Nemotron Model Reasoning Challenge. Deterministic Python solvers for each puzzle category — gravity, unit conversion, numerals, ciphers, bit manipulation, equations — infer the hidden rule, emit a step-by-step trace, and distill it into a LoRA adapter for Nemotron-3-Nano-30B.
Verified reasoning, not vibes The solvers generate a chain-of-thought corpus and keep only solver-correct traces — then a rank-32 LoRA distills that reasoning into the base model, so no solver code is needed at inference. Coverage: 100% on cipher, numeral, and unit-conversion; 99.7% on gravity; ~76% on bit manipulation.
Solvers6 categories
Coverageup to 100%
AdapterLoRA · r32
Four stages — Solve → Generate → Distill → Submit. Solvers turn algorithmic puzzles into verified (prompt, trace) pairs; LoRA SFT (rank 32, completion-only masking) teaches the base model to reproduce the reasoning; the adapter is scored in vLLM by extracting the
\boxed{…} answer. The holdouts are bit manipulation (~76%, solved with DSL search + GF(2) affine synthesis) and equation, which composes a per-problem symbol substitution with a numeric rule from only a handful of examples.entity resolutionUBID match
KSPCBSri Lakshmi Traders
BESCOMS. L. Traders
Shop Estಶ್ರೀ ಲಕ್ಷ್ಮಿ ಟ್ರೇಡರ್ಸ್
unified business id
UBID·KA·9253
Kannada ↔ Roman
match0.94
sources3 → 1
KARMA — Karnataka Active Record & Master Aggregator
Probabilistic entity resolution across 40+ Karnataka government department databases. It assigns one Unified Business Identifier (UBID) per real business by linking records from KSPCB, BESCOM, Shop Establishment, Factories, and Labour systems — with no LLMs and a per-field audit trail on every match decision.
🏆 Winner — AI for Bharat 2026 Built with team Progsolve for the Government of Karnataka hackathon. Verified on 49,857 synthetic records → 9,253 cross-source UBIDs in ~80 seconds, 101/101 tests passing.
The moats are Karnataka-native: Kannada↔Roman transliteration, KIADB shared-plot disambiguation, BESCOM RR-number variants, and Shop Establishment honorific noise — baked into the normalizer and scorer, not bolted on. 63% of MSMEs are informal with no PAN/GSTIN, so there's no reliable join key; KARMA scores matches field-by-field and stays auditable.
traffic optimizerroute scoring
SCH
BUS
CAR
road sizewide route +18%
school timingavoid 8:20 peak
vehicle typebus priority lane
AITO — AI Traffic Optimizer
An AI-assisted routing system for city-scale traffic. Distributes vehicles across viable routes using road capacity, school-timing patterns, and vehicle-type priorities to prevent bottlenecks before they form.
Real-world target Built for Hyderabad's traffic patterns. Usable by commuters, delivery platforms, and urban planners.
AITO treats traffic as a prediction-and-allocation problem, not just a map problem. The model scores routes dynamically based on time-of-day and vehicle mix.
pre-incident detectioncost optimized
API
DB
Q
drift signal
FAULTLINE
A pre-incident cascade-failure detector for microservice architectures. Combines deterministic drift signals with Bedrock-powered root-cause hypotheses to flag problems before they propagate.
Engineering constraint Designed under a ~$120 cloud budget cap. Expensive model calls only fire on real anomalies; cheap math handles the baseline.
Cost-aware AI architecture: deterministic math does the first-pass filtering, and the LLM only runs when the signal is worth investigating. That's how you build AI on a real budget.
03 / Working Set
The stack I actually ship with.
AI / ML
- Python
- PyTorch, TensorFlow
- HuggingFace Transformers
- LoRA, PEFT fine-tuning
- Scikit-learn, XGBoost
- NumPy, Pandas
Backend
- FastAPI, Pydantic
- Django
- Node.js, Express
- MongoDB, SQL, SQLite
- Docker
- REST APIs
Cloud / Infra
- AWS Lambda
- DynamoDB
- API Gateway
- Amazon Bedrock
- HuggingFace Spaces
- Linux, Git
Frontend
- React
- Next.js
- Tailwind CSS
- EJS, Bootstrap
- Vanilla JS
- Responsive UI
04 / How I Build
From a rough idea to a shipped system — the loop.
01
Frame
Identify the real user, the sharpest constraint, and the smallest useful version. Cut scope before writing code.
researchscopingconstraints
02
Architect
Design for the constraint — cost, latency, or safety. Choose the stack that matches the real problem, not the trendy one.
system designcost modelapi shape
03
Prototype
Build the full loop quickly: input, logic, state, output. One sprint to get the whole thing running, then prune ruthlessly.
fastapireactpytorch
04
Ship & Iterate
Deploy it where real users can find it. Instrument, measure, then iterate against data rather than against vibes.
dockerawshf spaces
05 / Lab Notes
Research threads I'm pulling on right now.
AI Agents With Real Guardrails
Most agent demos skip the part where a tool call has consequences. I'm building security and authorization layers that reason about where an instruction came from before letting an agent act — so prompt injection can't quietly drive a high-risk action.
Cost-Aware AI Architecture
LLM calls are the new cloud bill. I'm obsessed with patterns that keep AI systems affordable: deterministic pre-filtering, cache hierarchies, and knowing when not to call the model.
Distilling Verified Reasoning
Coming out of the Nemotron work: for deterministic tasks, the win is generating provably-correct chain-of-thought with solvers and distilling that into a LoRA. Verified traces beat scraped ones — the model learns reasoning that's actually right.
AIMA + Applied Work
Working through Russell & Norvig's AIMA alongside shipped projects. Every chapter gets tied back to a real build so the theory doesn't stay decorative.
06 / Second Brain
A local knowledge base that knows my work.
This runs entirely in your browser — no API keys, no backend, no data leaving your device. It handles typos and loose phrasing, and picks from several phrasings each time so conversations feel human rather than canned.
Ask about any project in depth, the stack, hiring terms, achievements, or anything else on this site. If something isn't covered, it'll say so and point you to me.
second brainLOCAL · PRIVATE
07 / Why Hire Me
What you actually get when you hire me.
01
End-to-end AI delivery
I don't hand off at layer boundaries. Model, API, frontend, infra — I ship the full system. Teams hire two engineers for this; I do it in one seat.
02
Constraint-first thinking
I've shipped under a $120 cloud budget, 6-week hackathon deadlines, and safety-critical rollback requirements. I design for the constraint, not around it.
03
Production mindset
Top 300 globally in the AWS 10,000 AI Ideas Challenge wasn't for a prototype — it was for a deployable idea. I think in terms of what gets to users, not what looks clever in a notebook.
04
Fast, readable code
My strongest habit is shipping a working version inside the first sprint. Clean enough to review, documented enough to hand off, simple enough to debug at 2 AM.
08 / Contact
Have an AI problem?
Let's solve it.
Open to AI/ML internships, full-stack engineering roles, freelance builds, and collaborations. Response within 24 hours.