Alex Lavaee

Applied AI @MSFTResearch | Building the Future of DevTools | Prev AI Research @harvardmed + @BU_CDS | AI Startups

[Alex Lavaee @alexlavaee]

Hi, I'm Alex! I'm a research engineer at Microsoft Research working on agentic engineering.

My hobby is contributing to open-source. My latest project is Atomic, an OSS verifiable coding agent runtime.

In the past I've built coding agents for a 1B LoC codebase, an FDA Deep Learning algorithm for aortic stenosis detection at Edwards Lifesciences, uncertainty estimation for 3D scene understanding at Themis AI (MIT CSAIL spinoff), and computer vision for satellite imagery at Boeing.

On the research side, I've post-trained self-supervised models for 3D human motion understanding at Boston University and spatial intelligence models for biological understanding (SpatialDINO @ Harvard Medical School).

Introducing myself and doing an AMA here to share and meet new people.

[Alex Lavaee @alexlavaee]

The game has indeed just started - we've only scratched the surface of possibility.

[Norin @norlava]

"I don't think of falling behind as much when I know this is the tool for the next generation of software development and the game has just started." - Sina, a senior engineer at GRiDD, whose GNet platform connects 6,000+ transportation companies worldwide, on Atomic (bastani-inc/atomic on Github), OSS verifiable coding agent runtime.

[Norin @norlava]

"I don't think of falling behind as much when I know this is the tool for the next generation of software development and the game has just started." - Sina, a senior engineer at GRiDD, whose GNet platform connects 6,000+ transportation companies worldwide, on Atomic (bastani-inc/atomic on Github), OSS verifiable coding agent runtime.

[Alex Lavaee @alexlavaee]

All open source via Atomic, bastani-inc/atomic on GitHub, verifiable coding agent runtime.

[Alex Lavaee @alexlavaee]

A verifiable coding agent runtime needs to make decisions backed by evidence.

Each handoff needs acceptance criteria and evidence that stages can inspect. Otherwise, as you run more agents you just are given more to review manually.

Sina, a senior engineer at GRiDD, whose GNet platform connects 6,000+ transportation companies worldwide, shares how verifiable workflows have become something he debugs and delegates work to. It enables him to reduce the evaluation bottleneck, ship with more assurance, and reduce context switching.

The opportunity is in combining dynamic reasoning capabilities of LMs and repeatable deterministic checks with advanced verifiers/reducers, not running more agents in parallel.

[Alex Lavaee @alexlavaee]

Good point, they are separate objectives. The routing is one piece of the puzzle where Atomic only moves to other candidates on failures. There is a separate mechanism in the workflow runtime itself that allows for rejecting completed runs via an LM as a verifier. And you can imagine that this can get complement the model router in the next iteration since the task description changes from previous iteration feedback.

[Alex Lavaee @alexlavaee]

We tried Jev to cut costs in long-running verification workflows by building a production-grade model router for the Atomic coding agent runtime. Routing reliably across hundreds of models while maintaining accuracy in long-horizon tasks was harder than expected. Hard-fought lessons 👇

[Alex Lavaee @alexlavaee]

Atomic is open source: https://t.co/D2l4vkapE0

[Alex Lavaee @alexlavaee]

One real Atomic run: explore → gpt-6-luna, minimal ($0.002) plan → gpt-6-luna, minimal ($0.001) review → Opus 5.5, xhigh ($1.37) The expensive reviewer caught issues in the cheaper planner's work.

[Alex Lavaee @alexlavaee]

I’ve been using @namespacelabs for my CI runners for Windows, MacOS, and Linux after hearing about them. They help power faster CI for Atomic (bastani-inc/atomic on GitHub), verifiable coding agent runtime. If you’re shipping cross-platform with coding agents you know how important that is. Very grateful for their team’s partnership and willingness to sponsor open source.

[Alex Lavaee @alexlavaee]

All code OSS in Atomic, verifiable coding agent runtime bastani-inc/atomic on GitHub

[Alex Lavaee @alexlavaee]

Test + Verification engineering best practices for coding agents that have been successful for teams shared in the video below.

Lots of low hanging fruit ways to optimize and reduce review bottlenecks.

[Alex Lavaee @alexlavaee]

https://t.co/F1KZRSQ1xD

[Alex Lavaee @alexlavaee]

Parallel constrained decoding does well when the task needs parsing, classification, extraction, or schema filling.

Don't reach for it when the task needs complex reasoning, understanding, or free-form text as autoregressive models are better for these cases.

[Alex Lavaee @alexlavaee]

https://t.co/N0roaqHpS4

[Alex Lavaee @alexlavaee]

Jev explained in under 10 min: what it is, how parallel constrained decoding works, when to use it, and when not to.

Shoutout to @NielsRogge + @gregpr07 for the great diagram + demo used in the educational video.

[Niels Rogge @NielsRogge]

For anyone curious how Jev works, I made a visual explanation using @claudeai :)

This is based on the Qwen2.5-RLCD model which @harshagundal released on @huggingface

The idea is to replace autoregressive LLM generation by a single Transformer decoder (of a pre-trained LLM), which processes the context + JSON schema only once. The keys and values of those tokens are cached.

Next, for each field of the JSON schema, we:
1. pass its field suffix tokens through the Transformer decoder again (reusing the KV-cache)
2. obtain a final hidden state, which we pass through the language modeling head
3. we obtain scores, also called logits, for all tokens in the vocab of the LLM
4. we only look at the scores of the tokens we care about for the given field, and pass those through a softmax to obtain probabilities which sum to 1
5. we take the token with the highest probability.

The benefits of this are that:
1. it's fast (we don't need to generate the JSON schema token by token)
2. it's 100% valid JSON (we don't need to rely on the model to generate a valid schema)