16 private repos
·
268 contributors
·
17,937 commits
·
56K classes
·
1,033 data models
·
396 endpoints
·
72.6% active test cov.
·
15 ontologies
·
3A 11B 1C
·
29 papers cited
Due Diligence at a Glance
The eight questions every technical DD asks, answered fleet-wide. Green passes; amber marks the items remediation is priced into.
Active Test Coverage
72.6%
13 of 16 repos above 60% · hand-written code only
Tests Run in CI
14 / 16
GitHub Actions 8 · Bitbucket 7 · Travis 1
Docker Builds
16 / 16
Every repo ships a Dockerfile · 10 build fully offline
Dependencies Pinned
13 / 16
Lockfiles in only 4 of 16 — the top remediation item
External Services
6 systems
AWS · MongoDB · Redis · RabbitMQ · Kafka · Elasticsearch
Contributors
268
17,937 commits · median bus factor 5
Confirmed Security Findings
20
Across 1.26M LOC after false-positive triage · fleet security 96.4 / 100
Quality Grades
3A · 11B · 1C
Fleet average 74.1 / 100 · deterministic 9-dimension scorer
Executive Overview
Technical due diligence for ForwardLane, an AI wealth-management platform. The asset: 16 private repos · 1.26M LOC (~8.8M tokens) · 268 contributors · 17,937 commits — 88% Python, with React/Node front-ends.
Every repo is scored on 9 quality dimensions and mapped to financial-industry ontologies — a labeled, domain-specific dataset that does not exist in public training corpora.
About ForwardLane
ForwardLane applies AI to wealth management: it ingests financial documents, extracts meaning (entities, fund data, signals, sentiment), ranks it by relevance to each advisor's clients, and delivers personalized recommendations.
Advisors use a React dashboard and a conversational client (Node.js); behind them, 14 Python services handle NLP extraction, sentiment, document ranking, portfolio analysis, and signal generation.
The Private Python Codebase as an AI Training Asset
TL;DR — 62K labeled function–concept pairs, SWE-bench tasks from real commits, and a 549K-edge call graph: private financial training data that public corpora don't have.
ForwardLane's codebase is a private, domain-specific dataset that does not exist in public training corpora. It contains real-world financial AI logic built by professional developers over years of production use. This makes it valuable for AI training in ways that public GitHub code cannot match:
RLHF & Fine-Tuning Data
- 62,453 function-to-concept pairs mapped against the Financial Industry Business Ontology (FIBO)
- Each pair is a labeled training example: "this function handles Client operations," "this function computes a RiskAssessment"
- Enables RLHF and supervised fine-tuning for financial domain LLMs
- Private data — does not exist in public training corpora
SWE-Bench Verified Benchmark Data
- SWE-bench format task instances generated from real bug-fix commits
- Each instance includes: bug description, before/after code, and verifying test
- Standard format used to evaluate AI coding agents
- Financial-domain specific — underrepresented in existing benchmarks
Agentic AI Navigation Data
- 549,772-edge call graph mapping every function to every other function
- Combined with ontology labels, enables AI agents to reason: "to change risk assessment, modify these 76 functions across 3 repos"
- Structured graph data for training autonomous coding agents
Why this matters: Public code repositories are already saturated in training data. Private codebases with professional-grade financial domain logic, labeled with industry-standard ontologies, and verified with SWE-bench format benchmarks represent a scarce and high-value data asset for training the next generation of AI coding models and agents.
Codebase Repository Overview
The codebase tells the story of how financial data flows through ForwardLane: raw documents come in, get processed and understood by AI, produce ranked investment insights, and reach advisors through a visual dashboard and conversational client.
Core Platform — Where the Business Logic Lives
forwardlane-backend
The heart of the platform. Manages client portfolios, ranks investment content for each advisor, ingests financial documents from data providers, and runs the AI recommendation pipeline. Contains modules for client ranking, portfolio analysis, content ingestion, document ranking, and user management. The largest and most business-critical repository.
A · 82.8
430K LOC
73.7% tested
430 models
76.8K functions
signal-studio-backend
The signal generation engine. Detects market signals, builds analytical dashboards, and feeds the investment recommendation pipeline. Shares a codebase heritage with forwardlane-backend and adds signal-specific analytical capabilities, portfolio analytics, and reporting tools for wealth management teams.
B · 70.2
372K LOC
75.6% tested
436 models
81.4K functions
signal-builder-backend
Lets users build custom investment signals from individual components. Built with clean architecture principles (FastAPI), this service handles signal composition, backtesting against historical data, and exposing composed signals through a well-documented API layer.
B · 72.4
55K LOC
77.2% tested
39 models
core-top-recommendations
The client-facing recommendation engine. Takes the output of all the upstream analysis — rankings, signals, sentiment — and surfaces the top investment actions for each individual client based on their portfolio, risk profile, and investment goals.
B · 69.2
7.2K LOC
83.5% tested
Data Intelligence — Understanding Financial Documents
fl_data_microservice
The NLP workhorse. Takes raw financial documents and turns them into structured data using natural language processing, entity extraction, sentiment scoring, and Elasticsearch indexing. Processes text using spaCy, NLTK, TextBlob, and Apache Tika.
B · 71.8
34K active LOC
62.2% tested
core-entityextraction
Identifies named entities — companies, funds, market indices, financial instruments — from unstructured financial text. Provides a REST API for entity recognition used by the broader platform.
B · 71.0
6.7K LOC
80.3% tested
entity-extraction-morningstar
Specialized connector for Morningstar data. Extracts structured fund, ETF, and market data from Morningstar's feeds and normalizes it for use across the platform.
B · 76.0
1.7K LOC
79.3% tested
core-document-ranking-manager
Determines which financial documents matter most for each client. Ranks and prioritizes content by relevance to individual portfolios, ensuring advisors see the most impactful information first.
B · 70.0
12.5K LOC
67.7% tested
26 models
core-sentiment-analysis
Trained machine learning models that classify financial text as positive, negative, or neutral. Provides a sentiment signal that feeds into the recommendation engine and client-facing insights.
A · 80.1
675 LOC
67.9% tested
core-scraping
Collects financial data from external sources via AWS, Kafka, and MongoDB. The entry point where raw market data, news, and research reports first enter the platform.
B · 72.9
9.6K LOC
74.3% tested
Machine Learning & Data Science
mars (Quantitative Research Platform)
ForwardLane's quantitative research and model development platform. Contains ML model training, evaluation, and deployment code for financial predictions and portfolio analytics. The highest-quality repo in the fleet with excellent documentation and type safety. Not related to the MARS paper or model.
A · 88.2
9.1K LOC
50.2% tested
django-etl-sync
A purpose-built ETL (Extract-Transform-Load) framework on Django. Synchronizes external financial data sources into the platform database. Has the highest test coverage in the fleet at 88.4%, reflecting its role as critical data infrastructure.
B · 68.8
10.1K LOC
88.4% tested
core-pipeline-manager
Orchestrates all the data pipeline jobs. Manages scheduling, task dependencies, retries, and monitoring across the entire data processing workflow from ingestion to output.
B · 73.7
3.4K LOC
78.5% tested
User-Facing Applications — How People Use ForwardLane
core-admin-ui
The React-based administration dashboard. Provides a visual interface for internal teams to manage client accounts, review generated signals, configure pipeline settings, and monitor system health.
React / JS
45K LOC
24.6% tested
forwardlane_advisor
The advisor-facing client application. A Node.js conversational interface where financial advisors interact with the platform, ask questions about their clients, and receive personalized investment recommendations. Contains 193 API endpoints.
C · 44.8
Node.js
561K LOC
3.6% tested
core-admin
The backend powering the admin dashboard. Provides APIs for user management, pipeline configuration, system monitoring, and platform administration. Connects to MongoDB and RabbitMQ.
B · 70.4
12.5K LOC
78.4% tested
Archipelago & APEX-Agents Integration
TL;DR — the knowledge graph, scores, and SWE tasks plug into the Archipelago/APEX evaluation stack as MCP tools, verifiers, and 16 new financial worlds.
Archipelago is an open-source harness for running and evaluating AI agents against MCP-enabled environments. APEX-Agents uses it to test whether agents can execute long-horizon professional tasks across 33 simulated "worlds" (investment banking, consulting, law) with 480 tasks. The top model scores just 24% Pass@1 — agents fail most often on code execution, document retrieval, and tool orchestration. ForwardLane's infrastructure addresses each of these failure modes directly.
Knowledge Graph as an Archipelago MCP Server
- Archipelago exposes tools via MCP servers (filesystem, code execution, spreadsheets, calendar, mail)
- ForwardLane's Knowledge Graph MCP server plugs directly into any sandbox as an additional tool
- Agents get real-time access to the 500K+ edge graph — query functions, dependencies, and impact before executing code
- Directly addresses the #1 agent failure mode: blind code execution without codebase understanding
- Integration is a single configuration change in the Archipelago environment
Financial Domain Worlds for APEX
- APEX-Agents has 10 investment banking worlds (172 files avg)
- ForwardLane can supply 16 additional financial SW engineering worlds
- Real production repos: client-ranking, portfolio analysis, NLP entity extraction, sentiment scoring
- Built by 268 professional developers — not synthetic
Each world comes pre-packaged with:
Analytical DB
FIBO labels
SWE-bench tasks
Call graph
Security findings
9-dim scores
fl-repos Scores as Archipelago Verifiers
- Archipelago uses binary pass/fail verifiers with weighted scoring
- fl-repos 9 dimensions map directly to this format
- Each dimension → verifier with configurable thresholds
- Gives Archipelago a code-quality grading layer it doesn't currently have
- APEX evaluates task outputs (spreadsheets, slides); fl-repos evaluates the code itself
APEX-SWE & SWE-Bench Alignment
- APEX-SWE evaluates SW engineering agents via Docker + MCP tools
- ForwardLane provides financial-domain SWE tasks in exactly the expected format
- Each instance: problem statement, source code, Docker env, verifying tests
- Fills a gap: existing APEX-SWE tasks are from open-source repos only
- Tests reasoning about client-ranking, portfolio calcs, entity extraction
Financial Evaluation Harness
TL;DR — fl-repos scores and FIBO labels supply ready-made ground truth for all four fin-eval dimensions.
The fin-eval-harness is a financial visual reasoning evaluation system with weighted scoring across four dimensions. ForwardLane's labeled codebase and knowledge graph provide the ground truth data this harness needs.
50%
FUNCTIONAL
Does the code work correctly? fl-repos Test Quality score (100/100 fleet avg) and SWE-bench instances provide functional verification.
25%
TRAJECTORY
Did the agent take the right path? The 549K-edge call graph defines optimal traversal paths through the codebase for any given task.
15%
ROBUSTNESS
Is it secure and reliable? fl-repos Security dimension (96.4% fleet avg) and CPG vulnerability detection map directly.
Is the code well-written? fl-repos Documentation (66.1% avg) and Type Safety (39.4% avg) dimensions evaluate exactly this. The automated flywheel actively improves these scores, generating docstrings, type annotations, and refactoring complex functions.
The integration path: fin-eval-harness uses an LLM grader (Claude Sonnet) that receives the question, ground truth, and model answer and returns CORRECT/INCORRECT. ForwardLane's 62,453 FIBO-labeled function-concept pairs serve as ground truth for financial reasoning tasks. The 1,033 data models provide structured schemas for financial visual reasoning. And the analytical engine enables instant ground truth queries for any evaluation dimension.
Strategic Value for AI Labs
TL;DR — training data, evaluation infrastructure, and programmable DD tooling, all generated by one repeatable pipeline.
ForwardLane's codebase and intelligence platform address three needs in an AI lab's ecosystem: high-quality training data for AI model improvement, structured evaluation infrastructure that parallels APEX's rubric methodology, and scalable talent assessment tooling that turns code analysis from a manual process into programmable infrastructure.
RLHF & Fine-Tuning Data
- 62,453 labeled (function, concept) pairs across 15 ontologies
- Pre-labeled, domain-specific data for supervised fine-tuning and RLHF
- Replaces manual expert labeling that domain evaluators would otherwise produce
- Private financial codebases — not available in public training corpora
Agentic AI Navigation
- 549K-edge knowledge graph accessible via MCP server
- Any AI agent (Claude, GPT, custom) can query it directly in real time
- Turns a static codebase into live, navigable infrastructure
- Agents can reason about impact before making changes
APEX-Aligned Scoring
- 9-dimension deep scorer follows APEX rubric philosophy
- Decomposes quality into discrete, objective criteria
- Every score traceable to specific functions — not subjective opinion
- Deterministic: run it twice, get the same scores
SWE-Bench Verified Data
- Auto-generated benchmark task instances from real bug-fix commits
- Standard SWE-bench format: problem, source code, verifying test
- Financial-domain specific — underrepresented in existing benchmarks
- Directly usable for APEX-SWE evaluations
Programmable DD Infrastructure
- Not a report — a repeatable pipeline
- Run against any codebase → structured scores, knowledge graphs, dashboards
- Replaces weeks of manual DD with 30-minute automated analysis
- Backed by proprietary analytical engine and knowledge graph
IP Mapping & Valuation
- FIBO mapping identifies where proprietary financial logic lives
- 1,996 functions handle Client operations
- 1,036 process Market Signals · 682 compute Investment Rankings
- Quantify domain-specific IP vs. generic infrastructure with data
How This Analysis Was Produced: fl-repos
TL;DR — every number on this page is generated by fl-repos, reproducible, and traceable to specific functions.
This report was generated by fl-repos, a proprietary code intelligence platform built by ForwardLane. Every number, score, and mapping in this document is reproducible and traceable to specific functions in the codebase.
For Talent Assessment
- Run against any candidate's codebase → 9-dimension quality score in minutes
- Replace subjective code reviews with quantitative, APEX-aligned rubrics
- Every score traces to specific functions — fully explainable, never a black box
For Due Diligence
- Evaluate acquisition targets, portfolio companies, or contractor codebases at scale
- 235K functions processed in under 30 minutes
- Interactive dashboards, knowledge graphs, and structured data exports
- Replaces weeks of manual review with a repeatable pipeline
For AI Agent Evaluation
- 500K+ edge knowledge graph accessible via MCP server
- Any AI agent (Claude, GPT, custom) can query it in real time
- Plug into Archipelago sandboxes for codebase understanding
- Directly addresses the #1 failure mode in APEX-Agents
Key Capabilities
9-Dimension Scoring
Documentation, testing, security, domain alignment, resilience. Deterministic — run twice, same scores.
15 Ontology Mappings
Finance, security, ML, provenance, cloud, data engineering. 62,453 labeled function-concept pairs.
AI Safety & Security
Context-aware vulnerability classification using call graph context, not pattern matching.
SWE-Bench Generation
Verified benchmark instances from real commit history in the standard evaluation format.
Improvement Flywheel
Closed-loop pipeline: +23 points (51→74.1) across 14 repos. 10 research innovations.
Zero-Trust, Three-Tier Architecture
1
Local Processing & Analysis
- All extraction, scoring, and analysis runs locally on your machine
- Source code never leaves the local environment
- High-performance embedded analytical database for instant queries
- Full pipeline: 1.26M LOC analyzed in ~8 minutes
2
GPU Cloud Fan-Out
- LLM-intensive tasks (ontology mapping, security audit, flywheel) fan out to cloud GPU containers
- 50-100 concurrent workers for massive parallelism
- Only function signatures and metadata leave the machine — not source code
- Full improvement flywheel: 45 min local → 5 min with GPU fan-out
3
Multi-Tenant Cloud Graph
- Results sync to a high-performance, multi-tenant cloud graph database
- 549K+ edges accessible via MCP server for real-time AI agent queries
- Tenant-isolated — each client's data is separated and access-controlled
- Enables Claude Code, ChatGPT, and custom agents to query the knowledge graph directly
Zero-trust data flow: Source code is analyzed locally and never uploaded. Only derived metadata (function signatures, scores, ontology mappings) flows to cloud tiers. The graph database is tenant-isolated with access-controlled API keys. End-to-end analysis completes in ~15 minutes with GPU acceleration or ~60 minutes fully local.
What Makes It Different
vs. Claude Code / Cursor
File-by-file tools. fl-repos analyzes entire codebases — cross-repo deps, call graphs, ontology maps. Quantitative, not conversational.
vs. SonarQube / CodeClimate
Lint rule checkers. fl-repos maps code to financial ontologies, AI training value, and business domains.
vs. Manual DD
5-10 files/hour manually. fl-repos: 235K functions in 30 minutes with interactive reports and knowledge graphs.
AI-Native Access
Knowledge graph MCP server exposes the full codebase to any AI assistant for direct querying. Runs locally, no cloud dependency.