Financial Services / Trading Technology · AI/ML Engineering · AgentCore POC · BUILDSTR
Challenge
Make every signal readable in seconds
Alstrum is the company behind Stryk, a trading platform that gives retail traders the kind of signals professional desks use. Its ML model, AlphaNet, scores every signal and attributes it with SHAP (SHapley Additive exPlanations) across 304 features, rolled up into 14 feature families. That attribution shows where institutional money is positioning before a move. But it surfaced as signed, ranked numbers, and Stryk’s Show Me and Guide Me traders needed to be able to read them.
- Every signal carried raw attribution values that only a data scientist could interpret
- Traders needed the why behind each call, in plain English, in seconds
- Accuracy came first: one invented or misranked driver would break the trust the product is built on
- A busy engineering team, shipping weekly, had no time to build and validate an explainer
- Per-signal cost and latency had to fit inside a credit-funded runway, with no disruption to live workloads
Solution
A faithful explainer on Bedrock AgentCore
In a focused four-week engagement, BUILDSTR built a serverless explainer layer in Alstrum’s own AWS account, beside live workloads, with no disruptions to the production environment. The layer takes the 14 family-level SHAP contributions and signal metadata (ticker, direction, confidence) and turns them into a “Why” pill: a short plain-English tag shown next to the signal in the app. BUILDSTR owned the build, evaluation, and documentation end to end. Alstrum’s input was a few hours a week of engineering time for weekly checkpoints and gold-set review.
- Pipeline: AWS Lambda sends each signal’s attribution to Amazon Bedrock AgentCore, which generates the pill. SHAP data never leaves Alstrum’s account.
- Model bake-off: Claude Haiku, Amazon Nova Lite, and Llama 3.1 ran side by side on the same prompts and data. They were scored on fidelity, cost per signal, and latency, with annual run-rate projections at the high and low end.
- Fidelity guardrails: Every pill passes two automated checks before it’s accepted:
- Every number the model writes must match the input exactly.
- The named market drivers, their ranking, and the tone must match the actual top SHAP contributors.
- Evaluation: Every input/output pair is stored and scored against a gold set agreed with Alstrum’s team. Mismatches are flagged for prompt iteration.
- Operations: Results are stored in Amazon DynamoDB and monitored in Amazon CloudWatch. Everything is delivered as infrastructure as code, with a knowledge-transfer session at handover.
Result
The missing piece of the product
The project ended with an evidence-backed verdict: the explainer is production-viable on Bedrock, and the best model for it is Claude Haiku. Alstrum chose Haiku for its Why pills and found that the explainer was the missing piece of the product, ranking it ahead of live market data.
- Each Why pill costs under a cent and returns in under a second, a 200% performance improvement
- Every explanation is automatically checked against the source data, so traders see no invented drivers
- Output is plain English with no raw numbers, written for a new trader
- Alstrum got a working pipeline as IaC, ready to integrate into Stryk, plus model benchmarks to plan costs as it scales
