Hi, I'm Abheesht.
SoftwareengineerbuildingattheintersectionofreliablesystemsandappliedAI.
I like building reliable, scalable software that makes people's lives easier.
MS Computer Science · Arizona State University
01 / about
A bit about me
I'm a software engineer who's worked across systems, backends, and applied AI.
I've built low-level systems software at Samsung Semiconductor, real-time backends at a European SaaS startup, and agentic AI pipelines for enterprise data infrastructure. Earlier on, I did IoT research that led to two IEEE publications.
I recently finished my MS in Computer Science at Arizona State, and I'm currently based in San Francisco.
There's a lot going on outside the terminal window. I take that part just as seriously.
education
MS Computer Science · Arizona State University
GPA 4.0 · May 2026
BTech Information Technology · Manipal Institute of Technology
2019 – 2023
based in
San Francisco, CA
02 / projects
Things I've built
Agent-Techs AI Pipeline
2025Teaching a swarm of AI agents to clean up the messy, mismatched data that every enterprise quietly drowns in. FAISS-powered entity matching, orchestrated on GCP Cloud Run.
Text2SQL
2024Getting a small language model to write SQL that's actually correct, and refuse the queries that would wreck your database. A knowledge-graph validation layer pushed it to 76.2% on Spider.
Scalable Graph Pipeline
2024Graph data is powerful but a pain to feed at scale. A real-time pipeline that streams Kafka into Neo4j on Kubernetes and keeps the graph fresh without choking on its own ingestion.
Samsung Diagnostic Tool
2023The unglamorous tooling that validates the TCON chips behind Samsung's displays before they ship. Built in C++ and PyQt across I2C/SPI, deployed to 3+ chip variants.
other work
Blockchain-based agricultural supply chain tracker. Smart contracts for traceability from farm to shelf.
Real-time application context monitoring with alerting and dashboard visualization.
Time-series forecasting model for airline passenger volumes using classical ML methods.
Binary classification model to identify poisonous mushrooms from UCI dataset features.
My first real encounter with applied AI — building systems that actually shipped at a national science museum. Both papers came out of a summer at NCSM and ended up being the reason I went deeper into ML.
MusoAssist: An Interactive Virtual Bot for Museum Gallery Guidance
Humanoid chatbot deployed at NCSM Kolkata. Non-monotonic conversation chains, IoT-activated physical exhibits. 73% comprehension vs 78% with a human guide.
Low-Cost Crowd Counting for Museum Gallery Management
P2PNet CNN on existing surveillance cameras. Output drove a motorized spotlight to the most-crowded exhibit in real time. Raspberry Pi + ESP8266, no new hardware required.
03 / experience
Where I've been
Five roles across research, systems, full-stack, and applied AI. Click any card to read the full story.
04 / what i use
What I use
$ cat how-i-use-ai.md
# how i use ai
AI is embedded in how I build. Cursor and Claude Code are my primary development environment for architectural reasoning, multi-file refactors, and debugging — not just autocomplete. I use v0 for rapid component scaffolding and CodeRabbit for automated review, even on solo projects.
I also build AI directly into products. The chat widget on this site streams responses from an edge runtime, uses prompt caching to reduce token cost, applies Redis-based rate limiting, and is grounded in a curated knowledge base to prevent hallucinated facts about me. I chose a smaller, faster model deliberately because model selection is a latency, cost, and reliability decision.
At Agent-Techs, I built multi-agent orchestration and RAG pipelines using LangChain, FAISS, and dense embeddings for entity resolution across healthcare, finance, and supply-chain data, with human-in-the-loop checkpoints for regulated domains.
I treat AI reliability like system reliability: evals, observability, guardrails, prompt-injection defenses, and the assumption that models will sometimes be confidently wrong. The skill is not just prompting — it is knowing what good output looks like and catching failures before users do.
Get in touch
Let's talk
Have a role, a project, or just want to argue about system design? I'm all ears.
If you're hiring, the timing is good.