from dataclasses import dataclass
@dataclass(frozen=True)
class Engineer:
name: str = "Shiva Ganesh Talikota"
role: str = "Product Development Engineer @ matriXO"
base: str = "Hyderabad, IN (UTC+05:30)"
degree: str = "B.Tech CSE, AI & ML — KPRIT, 2022-2026"
ships: tuple = ("retrieval systems", "agentic pipelines", "Next.js products")
depth: tuple = ("Python", "TypeScript", "PyTorch", "Node.js")
open_to: str = "SWE / AI-platform / product-engineering roles, 2026"
def __str__(self) -> str:
return f"{self.name} — {self.role}"I work on the product side of applied AI: taking a model or a retrieval pipeline that works in a notebook and turning it into something with a URL, auth, error states, and a person on the other end using it. Most of my time goes to matriXO, an EdTech platform that maps what students actually learn in college against what roles actually ask for, and closes the gap with targeted recommendations.
The rest goes to retrieval systems. I care about the unglamorous parts — chunking strategy, whether the retriever surfaced the right passage, and whether the answer cites something a user can verify. A confident wrong answer is worse than no answer, especially in compliance and education.
I also run @stable.speaks (44K+), where I explain AI and career tooling to students. Teaching a concept to 44,000 people who will call you out is a fast way to find the holes in your own understanding.
Experiments, workshops & hackathon builds
| Project | What it is | Stack | Link |
|---|---|---|---|
| agentic-ai-workshop-avn | Teaching material from an Agentic AI workshop I ran — tool-calling loops built up from scratch | Python |
Repo |
| scanseqjs | Browser-side sequence scanning — kept the work off the server entirely | JavaScript |
Live |
| infinova-hackathon | Hackathon build, shipped inside the time box | JavaScript |
Live |
| matrixo-website-deployment | matriXO Web Dev Internship — technical evaluation R1 | TypeScript |
Live |
| THK-Website | Client site, spec to deploy | JavaScript |
Live |
| snake-game | Collision detection and a game loop, written for fun | JavaScript |
Play |
The retrieval path is where the engineering decisions live, so here it is end to end:
flowchart LR
A["MSME scheme docs<br/>PDF · circulars"] --> B["Chunk<br/>overlap preserves<br/>clause context"]
B --> C["Embed"]
C --> D[("Vector index")]
Q["User query<br/>'Am I eligible?'"] --> E["Embed query"]
E --> F["Top-k retrieve"]
D --> F
F --> G{"Relevant<br/>context?"}
G -->|no| H["Decline<br/>· no invented eligibility"]
G -->|yes| I["Generate<br/>grounded in retrieved text"]
I --> J["Answer + citations<br/>· user can verify source"]
style H fill:#7f1d1d,stroke:#ef4444,color:#fff
style J fill:#14532d,stroke:#22c55e,color:#fff
style D fill:#312e81,stroke:#818cf8,color:#fff
The branch at G is the whole point. Anything that ships an eligibility answer to a small business owner needs a path where it refuses, and the citation trail exists so the user never has to take the model's word for it.
Ship with these regularly
Working knowledge — used on real projects, still deepening
Concepts I work in
RAG · vector search & embeddings · agentic tool-calling loops · transformers & attention · recommender systems · computer vision pipelines · REST API design · CI/CD
| When | Role | What it involved |
|---|---|---|
| 2024 – Present | Product Development Engineer — matriXO, Hyderabad | Own the product surface end to end: architecture, Next.js/Node build-out, the Python recommendation service, and release. |
| 2025 | Core Team — CSR Summit 2025 | Operations for a 1000+ attendee summit — logistics, scheduling, on-day execution. |
| 2024 | Speaker — T-Hub Hyderabad | Talk on EdTech systems to 200+ founders and students. |
| 2023 – Present | Creator — @stable.speaks | 44.1K+ following. AI tooling and career engineering, explained for students. |
| — | Intern — Student Tribe · Ex-TurboHire · Crowdsource by Google India | Product and data work across hiring tech and open data collection. |
| 2022 – 2026 | B.Tech CSE (AI & ML) — KPRIT, Hyderabad | Coursework in ML, DL, DSA, systems. |
Also building OSSEB.
Things currently open on my desk, not a curriculum:
- Evaluating retrieval, not vibes — building a labelled question set for pAIr so retrieval quality is a number that moves, instead of a judgement call after five manual queries.
- Agentic loops that fail well — tool-calling agents are easy to demo and hard to make recover from a bad tool response. Working on retry and fallback behaviour that does not spiral.
- Serving models cheaply — batching, caching and quantisation, because inference cost is a product constraint for a student-facing platform, not an implementation detail.
- Reading — the attention and retrieval literature, and other people's production post-mortems, which teach more per page than most papers.
Roles — Product Development Engineer · Software Engineer (Full-Stack / AI-Platform) · Applied ML Engineer
Available — internships and part-time now · full-time from mid-2026 (graduating B.Tech CSE, AI & ML)
Location — Hyderabad · remote · open to relocating within India
What I bring on day one — shipping Next.js and Node products, Python services, retrieval systems that are honest about what they do not know, and the habit of finishing things and putting them in front of users.
If you are hiring for any of that, or you want a second pair of hands on an AI product, email me or reach me on LinkedIn. I reply to everything.
Certifications
About this account
My earlier account @shivaganesht — which held 55+ repositories — was suspended without a stated reason, and I have an open case with GitHub Support to restore it. Active work lives here in the meantime, and older projects are being migrated across. If you have a route to escalate it, I would genuinely appreciate the help: shivaganesht@icloud.com.
