AVANT GARDE LABS Discuss a project

AI development & deployment

AI systems for
your business
and your customers.

I build AI applications, integrate them with existing systems, and deploy them in customer environments. You work directly with me, from the first technical discussion through delivery.

Explore the work Have a project in mind?

01 / Working together

Project or contract

A

AI development

LLM applications, computer vision, and workflow automation. I handle the application and the integrations it needs to be useful.

B

Fixed-scope builds

A defined deliverable, agreed scope, and fixed price. Suitable for a pilot, an integration, or a complete application with clear requirements.

C

Embedded / FDE

I join your engineering team or work directly with your customers to implement and deploy your product. Available on a contract or retainer basis.

ERP modernization

Dashboards, integrations, and automation around the systems already in use. The Tally work below shows how this can be done without a migration.

02 / Selected work

Systems & source code

Examples of the applications, integrations, and deployment work I take on.

01 — Multimodal AI / workflow automation

QuoteAssist

In production

From a customer’s message to a priced quotation.

Accepts text, voice notes, and photos of handwritten lists. It combines the request with pricing data to generate a quotation or proforma invoice.

  • Conversation context carries through follow-up requests.
  • Available on WhatsApp and Telegram, with an independent web interface.
  • LLM orchestration with structured outputs for downstream processing.
  • Price lists and LLM providers can be updated without restarting the service.
Discuss a similar build Product page
Demo / Message to quotation01:23
Engineering notes

The workflow brings messaging, OCR, speech transcription, pricing, and document generation into one application. Tally is connected through its XML API over a WebSocket bridge.

Application
Rust · Tokio · Axum
AI & input processing
Claude / Groq · Textract · Whisper
Messaging & storage
Twilio · Telegram · PostgreSQL
Deployment
Docker · DigitalOcean App Platform

02 — LLM classification / business workflows

QuoteWatch

In production

An ambient agent for quotation follow-ups.

Runs in ephemeral sandboxes, reads Gmail threads, and prepares two work queues: customers waiting for a quotation, and sent quotations that need a follow-up. It works alongside the existing sales workflow; the team keeps using its inbox without maintaining a separate CRM.

  • Structured LLM responses are validated before a run is committed.
  • Atomic writes and an idempotent run key prevent duplicate records on retry.
  • Existing Gmail labels provide explicit signals for closed or ignored work.
Read the source Product page
Demo / Inbox to report00:42
Engineering notes

A run is prepared and validated in memory, then committed in one database transaction. Incomplete classifier batches are retried per thread. The monitored Gmail account uses read-only access; message digests are sent to the configured LLM provider for classification.

Application
Rust · Gmail API
Classification
Structured output · Anthropic / Fireworks
Persistence
libSQL · Local or remote database
Verification
Regression scenarios · Retry & rollback tests
Architecture documentation

03 — Local inference / developer tooling

VisionGrep

Open source

Search images by what’s in them.

An agent-first image-search tool in Rust, usable by coding agents such as Codex and Claude Code through its command-line interface. Search with a description or a reference image. CLIP inference runs locally; SQLite caches image embeddings for subsequent searches.

  • Works offline after the required models are downloaded.
  • Supports JSON output and a persistent stdio interface.
  • Includes a benchmark harness for performance and behavior comparisons.
Read the source Benchmarks
visiongrep / usage

Search with a description

visiongrep "a fancy red hat" ./photos

Or use a reference image

visiongrep --image ./reference.jpg ./photos
Local images CLIP embeddings / cached in SQLite Ranked image paths

macOS & Linux · Rust

Engineering notes

Unchanged images reuse their cached embeddings. The persistent stdio mode keeps loaded models resident between requests and checks the directory for changes before each search. Similarity scores are cosine similarities, not probabilities.

The repository documents setup, benchmarking, and current limitations, including concurrent access during reindexing.

Inference
Local CLIP text & image models
Index
SQLite · Incremental embedding cache
Interfaces
CLI · JSON · Persistent stdio
Evaluation
Latency · Resource use · Ranking & behavior

04 — ERP modernization / inventory analysis

StockIQ

In production

Inventory analysis alongside an existing Tally installation.

A browser-based dashboard for cable and wire distributors. It reads Tally’s built-in XML API and runs on the same Windows machine, without a plugin or data migration.

  • Cut-length analysis, reorder planning, and a stock grid by brand and size.
  • Per-item movement, turnover, margin, and inventory history.
  • Rust analysis compiled to WebAssembly; usable across the local network.
Read the source Product page
Demo / Inventory analysis00:35
Engineering notes & stock grid

An Axum server queries Tally over HTTP/XML. The browser runs the Rust analysis engine through WebAssembly. Deployment is a single executable and a static folder; operation does not require an internet connection.

Server
Rust · Axum · Tokio
Analysis
WebAssembly · wasm-bindgen
Data source
Tally ERP XML API
Frontend
HTML · JavaScript · CSS
StockIQ wire stock grid, with stock quantities organized by brand, wire size, and colour
Wire stock by brand, size, and colour. Select the image to inspect it at full size.
Tally XML parser Tally client

03 / Background

The person behind the work

Uday
Jhunjhunwala.

Engineer & founder
Avant Garde Labs

I was an engineer at Microsoft and Amazon, built two startups, and spent more than a decade running an electrical distribution business.

Several of the systems here began as tools for that business. Working with the people using them shaped how I approach software: understand the workflow, account for the existing systems, and follow through on deployment.

My background also includes production Rust infrastructure and Solidity and Cairo smart contracts on Ethereum and Starknet.

More work on GitHub

How an engagement starts

  1. 01

    Understand the work

    We discuss the workflow, existing environment, and what you need delivered.

  2. 02

    Agree on the engagement

    A fixed scope for a defined build, or a contract for work within your team.

  3. 03

    Build, deploy, hand over

    Regular progress reviews, deployment in the agreed environment, and documentation for the team running it.

04 / Contact

Start a conversation

What are you
working on?

Tell me what you’re building, what’s already in place, and where you need help. An outline of the scope and timeline is a useful starting point.

automation@avantgardelabs.in

AI development · Fixed-scope builds · Embedded / FDE contracts