~/projects — zsh

Last login: Sun Oct 18 07:00:53 2026

trisham@portfolio:~$ whoami

→ AI engineer building LLM systems, agents and production backends

trisham@portfolio:~$ cat headline.txt

AI Engineering and Full-Stack Projects

trisham@portfolio:~$ cat about.txt

// Voice agents, multi-agent systems and LLM routers, plus a

// from-scratch EAGLE-3 reproduction for faster LLM inference

// and full-stack sites running in production.

trisham@portfolio:~$ ./explore

online · 7 projects · 2 areas zsh · 120x40

# 01 — live_products

Live products,
shipped to real customers.

trisham@portfolio:~$ cat dhokla-house.md

dhokla-house.md

Dhokla House

// Production food-ordering and brand platform built for menu discovery and fast checkout.

Frontend: React.js with modular UI components and responsive layouts, so browsing, add-to-cart and checkout stay fast on mobile and desktop.

Backend: Node.js and Express.js APIs for menu retrieval, customer interaction flows and transactional order endpoints, with route and service layers kept apart.

Data layer: PostgreSQL holds the order lifecycle and reporting data. MongoDB stores content and interaction metadata whose schema changes often.

  • 2022–2023
  • React.js
  • Node.js
  • Express.js
  • PostgreSQL
  • MongoDB

visit_site

trisham@portfolio:~$ cat apni-dukaan.md

apni-dukaan.md

Apni Dukaan

// E-commerce platform covering catalog browsing, checkout and order management.

Frontend: Reusable catalog, product-detail and checkout components in React.js, designed to remove friction from the purchase flow.

Backend: Express APIs for products, cart and orders, with read-heavy catalog routes separated from transactional checkout and order-update routes.

Data and scaling: PostgreSQL for consistency-sensitive orders and reporting, MongoDB for flexible product metadata, so each side scales as volume grows.

  • 2022–2023
  • React.js
  • Node.js
  • Express.js
  • PostgreSQL
  • MongoDB

visit_site

# 02 — ai_and_llm_engineering

AI and LLM engineering,
from research to running systems.

trisham@portfolio:~$ cat thermacore.md

thermacore.md demo · 0:22

ThermaCore

// AI for data center operations, built for the AMD ACT III Hackathon.

ThermaCore launch video: the 3D data center digital twin on simulated telemetry.

What it does: Ingests server telemetry through Kafka, assesses each server's thermal and cooling state with deterministic rules, and forecasts inlet temperature with XGBoost.

Agents: Operators investigate through a LangGraph multi-agent recommendation workflow, a natural-language copilot and a digital-twin API built for a Three.js frontend. The system only recommends; it never controls hardware.

  • All telemetry is simulated from a synthetic Kaggle dataset and labelled as such.
  • Python
  • Apache Kafka
  • XGBoost
  • LangGraph
  • DuckDB
  • FastAPI

view_on_github

trisham@portfolio:~$ cat autopay-recovery-voice-agent.md

autopay-recovery-voice-agent.md demo · 0:23

Autopay Recovery Voice Agent

// FastAPI backend for a simulated autopay and EMI payment-recovery voice agent.

Voice agent demo: the LLM drafts each reply, and the backend blocks any claim it cannot verify, such as confirming an unconfirmed payment.

What it does: A virtual clock and event ledger drive payment and recovery state machines. Recovery calls go through an approval queue, and no call is placed without explicit human approval.

Voice and data: Twilio webhooks, a conversation engine with transcripts and audit logs, and a read-only natural-language-to-SQL tool. All data is synthetic and payments are simulated.

  • Python
  • FastAPI
  • SQLAlchemy
  • SQLite
  • Twilio
  • Fireworks AI

view_on_github

trisham@portfolio:~$ cat llm-router.md

llm-router.md

LLM Router

// Hybrid neural network and XGBoost router that picks a Claude model version for each query.

What it does: Sends every prompt to the cheapest Claude model that can handle it, trading off predicted quality, latency and cost. Prompts are embedded, classified as easy, medium or hard, and routed to Haiku, Sonnet or Opus.

  • Python
  • PyTorch
  • XGBoost
  • scikit-learn
  • Fireworks AI

view_on_github model_card

trisham@portfolio:~$ cat eagle-3-reproduction.md

eagle-3-reproduction.md

EAGLE-3 Reproduction

// Runnable PyTorch reconstruction of EAGLE-3 speculative decoding for faster LLM inference.

What it does: Rebuilds the EAGLE-3 draft model, multi-layer feature fusion, training-time test and lossless verification, with a chain-draft inference engine and one script per correctness experiment.

  • Greedy EAGLE-3 output matches the target model token for token.
  • Not the official implementation, and it does not measure the paper's GPU speedups.
  • Python
  • PyTorch
  • Speculative decoding

view_on_github read_the_write-up

trisham@portfolio:~$ cat acme-onboard.md

acme-onboard.md

ACME Onboard

// AI employee-onboarding platform with RAG over company documents, built for the IBM Bob 2.0 Hackathon.

What it does: HR uploads onboarding documents by department through a management portal. Employees ask an AI knowledge assistant that answers with department-aware retrieval-augmented generation.

Retrieval: Hybrid search with BM25 and embeddings, fused with RRF, plus HyDE query expansion, MMR diversification and Qwen3 reranking.

  • Next.js
  • TypeScript
  • FastAPI
  • ChromaDB
  • RAG
  • Fireworks AI

view_on_github

trisham@portfolio:~$ ls ~/more

github hugging_face_models engineering_writeups