Peach Pilot - Full Stack Engineer

Senior Full-Stack Engineer Sales Optimization · NLP Predictive

Atlanta, GA (Buckhead) | On-Site / Hybrid | Engineering Team

Most AI companies sell tools. We deliver the outcome.

Peach Pilot is an AI-Native Services (AINS) startup. Instead of selling software and leaving the work to the customer, we become the service provider and use AI to win, faster, better, and cheaper than a traditional firm, and we get paid when the results land. The customer buys an outcome, not a seat, and our system gets smarter with every engagement.

We are starting with life insurance sales optimization: helping agencies sell more effectively and more durably by learning, from their own data, what actually separates good outcomes from bad ones, and putting that back in front of the humans doing the selling. Our first client engagement is live and funded.

Peach Pilot is co-founded by Mario Montag (founder of Predikto, acquired by a Fortune 50, and an alum of McKinsey and PwC) and JP James (Hive Financial Assets, Georgia Tech, TITAN 100). We have a working platform with live infrastructure and a proven data-to-insights methodology.

The Role

You will join our engineering team as a senior full-stack contributor, owning a core layer of the platform, from the pipelines that turn tens of thousands of recorded sales calls into structured, governed data, to the applications that make AI-generated insights legible and actionable for people who have never touched a terminal.

The near-term problem is concrete and large: process roughly 50,000 recorded sales calls, extract structured metadata from them with NLP, and build the causal and predictive models that reveal what drives good outcomes, so we can optimize how humans sell. Backfill and live share one codebase. The same pipeline that mines the historical corpus also processes call transcripts in real time and feeds them into applications that coach the human while the call is happening.

We are building agentic-first, evidence-first software. Because our recommendations change how real businesses operate, correctness matters. Everything that touches a model output is versioned and reproducible, and no prediction that drives a real-world decision ships without validation. If you have built this kind of system before and want to do it at a startup where your work directly moves a customer's numbers, this is it.

What You Will Own Build

Call Ingestion NLP Extraction Pipeline

Build the pipeline that ingests roughly 50,000 recorded calls and turns them into a validated, structured dataset. Transcription (ASR), then NLP and LLM-based extraction of the metadata that matters (entities, intents, objections and how they are handled, sentiment, script adherence, outcomes), each tied back to its evidence in the transcript. You own data quality (deduplication, enrichment, classification, validation) and the governance boundary: PII minimization at ingest and access control that is enforced, not assumed. Backfill runs in batch, and the same code runs live on new calls.

Analytics, Causation Prediction

Build the feature and modeling layer that turns extracted metadata into insight: the analytics that identify which behaviors correlate with durable outcomes, the causal methods that separate correlation from a real lever (randomized holdouts and uplift modeling, not feature importance alone), and the predictive models (calibrated risk and opportunity scores) that let us act ahead of time. You will own the feature store, model registry, and the discipline that keeps a prediction defensible.

Live Transcript Processing in Applications

Incorporate real-time transcript processing into the product. Streaming, low-latency coaching delivered to the human on the call, with human-in-the-loop governance and confidence-tiered nudges (act firmly, suggest softly, or stay silent). You are responsible for how that guidance reaches the user and whether it earns trust or breaks their flow.

Client-Facing Dashboards Insight Delivery

Build the applications and interfaces that deliver the insights, the moment a client sees what is driving their results, backed by their own data, in plain language with rich visual context. These surfaces must be repeatable and deployable across every engagement through a per-client config layer. This is what sells the engagement and earns client trust.

The Stack

Frontend: React/Next.js, TypeScript, Tailwind CSS, Plotly, D3, Recharts

Backend: Python (FastAPI), Node.js/TypeScript

NLP AI/LLM: ASR/speech-to-text, NLP (NER, intent/objection classification, topic modeling, sentiment, embeddings and semantic search), Anthropic Claude, OpenAI GPT, LiteLLM (multi-model routing) for structured extraction with evidence spans

Analytics ML: Causal inference and experimentation (randomized holdouts, uplift and A/B, survival analysis), calibrated predictive modeling, feature store and model registry (MLOps)

Data Pipelines: BigQuery/DuckDB and object store, PostgreSQL, Redis, vector store,…