Aman Jha

AI Engineer — agents, retrieval & fine-tuning

AI Engineer focused on building production-grade LLM and agentic AI systems, with experience in long-term agent memory, retrieval pipelines, and domain-specific model fine-tuning. Experienced in taking AI systems from research and experimentation to deployable, production-ready services.

experience:

Agentic AI Engineer

Zuptek Infotech · Remote

January 2026 to July 2026
  • Built autonomous AI browser agents for the company's flagship product, a platform that tests whether AI agents can complete real tasks on business websites, using the OpenAI API, Playwright, and Chromium to read pages, interact with elements, fill forms, and complete end-to-end user journeys autonomously
  • Built the agent orchestration layer to plan and execute tasks step by step using structured outputs, with safety and reliability controls including permission limits, retries, time/step budgets, and network restrictions
  • Implemented prompt versioning with Jinja2 templates to systematically iterate on agent prompts, and built agent evaluation tooling combining deterministic checks with LLM-based evaluation to verify task completion, detect failures and loops, and benchmark agent performance
PythonOpenAI APIPlaywrightChromiumJinja2LLMs

Full Stack Developer Intern

Zuptek Infotech · Remote

July 2025 to January 2026
  • Built and shipped full-stack web products for clients using Next.js, React, and TypeScript, developing responsive interfaces with Tailwind CSS and turning product requirements into production-ready features
  • Developed backend functionality, APIs, and database integrations using Prisma and PostgreSQL, while debugging issues and iterating on products based on client requirements and feedback
Next.jsReactTypeScriptTailwind CSSPrismaPostgreSQL

projects:

MetaMemory

  • Persistent memory infrastructure for AI agents, scoring 89.0% task-averaged accuracy on LongMemEval-S (84.4% overall, 500-question long-term memory benchmark), evaluated with an LLM-as-judge Gemini reader
  • Retrieval pipeline combining semantic, temporal, BM25 keyword, and reranking channels, merged with Reciprocal Rank Fusion and re-weighted by a rule layer that learns from outcomes logged on prior queries
  • CABAR, a context-budget-aware retrieval pipeline for long-context agents: greedy token allocation, lexical MMR, recency weighting, and extractive summarization — cutting context usage by ~72-76% in internal evaluations at equal benchmark accuracy
TypeScriptPostgreSQL (pgvector)RedisNeo4jOpenAIGemini

DRT — Decomposed Relevance Tensors

  • Engineered and benchmarked an alternative retrieval-scoring architecture — splitting each embedding into k=6 weighted sub-vectors instead of one flat similarity score — against a cosine-similarity baseline trained on the full MS MARCO dataset (8.84M documents, 6,980 queries)
  • Ran a 4-step diagnostic (scoring ablations, per-slot logistic probes, failure analysis, linear CKA) to isolate why it underperformed baseline by 2% MRR@10, tracing the cause to encoder representation drift (CKA 0.945) rather than a broken scoring function
  • Open-sourced the full pipeline — training, evaluation, and diagnostics — with all raw results and logs checked in for reproducibility
PythonPyTorchMS MARCO

Sealevel

  • Coding LLM fine-tuned specifically for Solana/Anchor development: QLoRA fine-tune of Qwen2.5-Coder-7B on 270k Solana-specific records
  • 85% on an 80-task Solana/Anchor benchmark, 90% on general HumanEval
  • RAG pipeline over current Solana docs (Qdrant) keeps it accurate as Anchor syntax evolves
  • Served via an OpenAI-compatible API across web, CLI, VS Code, and MCP, with direct integration into Claude Code, Cursor, and Windsurf
  • 1,500+ installs on PyPI
PythonPyTorchSGLangQdrantMCPNext.js

Orky

  • Enterprise agentic AI platform that orchestrates workflows across 17 live app clients (ServiceNow, AWS, Jira, Salesforce, SAP, Snowflake, SharePoint) via a function-calling loop where Gemini selects and chains real API actions, with progress streamed over SSE
  • Enforces role-based access control at query time, filtering RAG results against each user's designation before content reaches the LLM, so employees only retrieve answers from HR, IT, CRM, and compliance sources they're authorized to see
  • Multi-step workflow engine that sequences AI agents with variable passing between steps, mid-workflow file-upload pausing, and automatic email notification on completion
FastAPINext.jsGoogle GeminipgvectorPostgreSQLAWS

Gifsy

  • Built a tool that turns any photo into a GIF, a sticker, or an interactive 3D scene, with all AI running in the browser using onnxruntime-web
  • Used Depth Anything V2 for depth maps, ISNet for background removal, and LaMa for inpainting the backdrop behind the subject
  • Wrote the 3D renderer in Three.js: depth-displaced planes with 7 motion presets, shared between the editor and the embed viewer
  • Split the depth model into a public encoder and a server-side head so free-tier generations are counted on the server and can't be bypassed from the client
  • Shipped auth, storage, embeds, and one-time Pro payments with Supabase, Vercel Blob, and Dodo Payments
Next.jsTypeScriptThree.jsonnxruntime-webSupabaseVercel Blob

Hodl.fun

  • Cross-chain token launchpad on Push Chain, built from zero to production — owning architecture, smart contracts, backend, and frontend
  • Solved a core DeFi UX problem (token launches trapped on a single chain) by enabling launches and trades with assets from any supported chain (ETH, SOL), no bridging required, via Push Chain's universal account system
  • Scaled to ~800K transactions and 500K+ token launches on testnet
  • Modular, upgradeable bonding-curve engine: separate Core orchestrator, Factory, per-token BondingCurve, and an ERC4626 FeeVault, with constant-product AMM math and virtual reserves for smoother price discovery
  • Integrated Uniswap V3 for automatic DEX listing once a token's bonding curve hits its liquidity target; all contracts upgradeable via UUPS proxy with reentrancy protection, access control, and invariant checks
  • Full-stack system: Foundry smart contracts, NestJS backend, Next.js frontend — Runner-Up, Project G.U.D hackathon (Push Chain)
SolidityFoundryTypeScriptNestJSNext.jsUniswap V3

achievements:

7 wins across grants and hackathons

Grants

  • Solana Foundation × CoinDCX India Grant: $10,000 for Lavinth (May 2025)
  • Solana Foundation × CoinDCX India Grant: $4,000 for tradeX (Feb 2025)
  • Dialect Instagrant: $1,000 for X-Mailer (Jan 2025)

Hackathons

  • Agents in Action Hackathon: 🥉 Akash Track (Best Overall Project), $2,000 Akash Credits, for PayPerUse
  • Agent Arena – University Hackathon: IQAI MVP Winner, for ResearchOS
  • Zo House Hackathon: Data Management AI Agent Track, 420 USDT, for DockMind
  • Project G.U.D (Push Chain): Runner-Up, for Hodl.fun

blog:

i build with:

PythonTypeScriptPyTorchQLoRA Fine-tuningRAG PipelinesOpenAIGoogle GeminiQdrantPineconepgvectorNeo4jRedisPostgreSQLFastAPIHonoNext.jsMCPSGLangDockerAWS

if you're hiring for AI/ML engineering, or want to build agents, retrieval systems, or fine-tuned models together, feel free to dm me on x or drop me an email.