Website:
crafter.co.in
Job details:
NOVA · AIRA
Machine Learning Engineer — Detection & Intervention Systems
Full-time · Hyderabad (on-site/hybrid) · Reports to Founder & CEO
Who we areNova is a bootstrapped, pre-seed startup building AIRA — an execution intelligence layer that sits inside Slack, Google Workspace, and Microsoft Teams. AIRA detects communication breakdowns inside a team before they become delivery failures, and does it quietly, without turning into another dashboard or another bot that gets muted.
We are six people, bxxased out of T-Hub in Hyderabad, part of the Google for Startups India Hub. No layers, no politics, no committee decisions — everyone here owns outcomes end to end.
The problem we're solvingMost teams don't fail because of one dramatic event. They fail slowly — a clarification loop that repeats three times, a decision that never quite lands, a thread that goes quiet at the worst moment. Jira tracks what was planned. HR tools measure how people feel, after the fact. Nobody sits in the middle, watching the actual conversation, and asking: is work breaking down here, or is a person breaking down here — or both?
AIRA is built on two pillars:
- Execution friction — clarification loops, stalled decisions, ownership gaps, blocked work.
- Emotional friction — disengagement, overload, tone shifts, boundary erosion.
Every signal AIRA produces resolves to three layers: the raw Signal observed, the Interpretation of what it likely means, and the Impact — why it matters for delivery or for the person. We lead every conversation with execution, because that's what gets us in the door. The emotional layer is what makes teams stay. This engineer will be building the machinery underneath both.
What you would actually be buildingThis is not a "plug an LLM into everything and see what happens" role. AIRA's detection system leads with fast, explainable, rule-based and NLP logic, and calls on machine learning selectively — only where deterministic logic genuinely runs out of precision. At a capability level, you'd be working across:
Capability area
What it means in practice
Signal detection
Building the systems that read team communication and detect patterns like repeated clarification, stalled decisions, and unclear ownership
Rule-based / NLP systems
Designing fast, explainable, deterministic detection logic as the primary approach — not defaulting to a model for everything
Selective ML / LLM use
Knowing when a problem genuinely needs a model versus when a rule or state machine is the better tool, and building the routing between them
Confidence & calibration
Scoring how sure the system is before it's allowed to act on what it detected
Intervention logic
Deciding, in a structured and tunable way, when the product should stay invisible, nudge gently, or escalate — and avoiding notification fatigue
Data & evaluation
Building and maintaining the datasets and test suites that keep detection accurate as the product evolves
A note on how we think about this, because it matters: state-tracking logic in our system is deterministic — it is code, not a trained model, and it should never be treated as one. Machine learning is called in selectively for reasoning and classification tasks that need it, not as a default fallback for logic that rules should be handling. If you're the kind of engineer whose first instinct is "let's just prompt a model for this," this role will frustrate you. If your instinct is "can a rule or a state machine do this more cheaply, faster, and more predictably — and where exactly does that stop working," you'll be at home here.
The full architecture — how the pipeline is structured, where exactly the boundaries between rule-based and ML-based logic sit — is something we walk through in the interview process, not in a public posting.
Responsibilities- Build and maintain rule-based and NLP detection systems as the primary decision-making layer — fast, explainable, and cheap to run at scale.
- Design the logic that decides when a signal is confidently resolved by rules versus when it genuinely needs deeper model-based reasoning.
- Own confidence scoring and thresholds — the layer that decides whether the product is even allowed to act on what it detected.
- Implement and tune a tiered intervention model — from invisible observation through to escalation — including cooldown logic so the product never nags.
- Build and maintain the datasets and test suites that keep detection accurate as the product evolves, and close the feedback loop from real usage back into the system.
- Work directly across major workplace communication platforms and APIs (chat, email, calendar, docs, and adjacent tools).
- Ship inside a Python/FastAPI + PostgreSQL backend, with a clear line to the dashboard consuming your outputs.
- Sit close enough to the founder and the rest of the six-person team that architecture decisions get made in a conversation, not a design-doc review cycle.
What you need to bringRequired
- Strong Python engineering — production code, not notebooks. FastAPI or an equivalent async framework.
- Real experience with NLP: rule-based / pattern-based detection systems, not just calling a model API and hoping.
- Comfort designing deterministic systems — state machines, threshold logic, scoring/calibration — where correctness and explainability matter as much as accuracy.
- Working knowledge of PostgreSQL and building/maintaining datasets for training, validation, and regression testing.
- Judgment about when to use an LLM and when not to. You should be able to argue against using one.
Strongly preferred
- Experience with Amazon Bedrock or equivalent managed LLM infrastructure, and prompt/response design for multi-step reasoning tasks.
- Experience integrating with Slack, Microsoft Teams, or Google Workspace APIs.
- Prior experience at an early-stage startup where you owned a system end to end, not a slice of one.
- Exposure to synthetic dataset generation for classification tasks.
The attitude that works here- You'd rather ship a simpler rule that works than a clever model that's a black box — but you know exactly when that trade-off flips.
- You're comfortable with ambiguity moving fast — specs here evolve as we learn, and you adapt without needing everything locked down first.
- You push back with reasoning, not ego, and you don't need to be told twice once a decision is made.
- You care about precision in language and architecture as much as in code — if a term is ambiguous, you'll ask before you build.
- You want ownership, not a ticket queue. In a six-person team, whatever you build, you also watch in production.
- You're not looking for a place that has it all figured out. We don't. You're looking for the chance to figure it out and get real equity in the outcome.
Company dynamics — what to actually expectWe are pre-seed, bootstrapped, six people, no revenue yet. We have one active pilot conversation in progress with an enterprise — not yet signed, not yet in production, and we're deliberately conservative about how we describe that, internally and externally. We are currently raising our first institutional round. None of this is a caveat we're hiding — it's the actual stage, and it means your work will directly shape whether this company has a next round, a next customer, and a next hire.
Decisions move fast and get revisited when we learn something new. Founder feedback tends to arrive as direct, numbered, specific instructions — not vague vibes. If you need six weeks of process to feel comfortable shipping, this isn't the right seat. If you want to be one of the first people who decided what AIRA's detection engine actually looks like, it is.
CompensationBase salary + meaningful early-team equity, calibrated to pre-seed stage. Exact figures discussed directly on a call once the screener below is cleared — we don't want to waste your time or ours with a mismatch on either side. [Insert current band before publishing.]
Before you apply — quick self-checkAnswer honestly. If any answer is "no," this likely isn't the right fit right now — please self-select out rather than apply anyway.
- Can you commit to working on-site or hybrid out of Hyderabad? (Yes/No)
- Have you built a production rule-based or NLP detection system before, not just called a third-party LLM API? (Yes/No)
- Are you comfortable joining a 6-person, pre-revenue, pre-seed team with no guaranteed runway beyond the current raise? (Yes/No)
- Can you reason through a rule-based/deterministic approach before defaulting to "just use a model"? (Yes/No)
- Are you available to start within [X weeks] of an offer? (Yes/No)
How to applySend your resume and a short note on one system you built where you had to decide between a rule-based approach and a model-based one — and why — to hiring@novacommunicate.com. Include links to code you're proud of if you have them. Skip the cover letter; the note above is the cover letter.
Click on Apply to know more.
