Jarvis

Your go to financial summarizer

Comment

Inspiration

India's retail investor base hit a massive 180M+ demat accounts this year. However, we noticed financial publishers are struggling with a huge content bottleneck. The demand for short-form video outpaces production capacity by 15-20x. Usually, a daily market update requires a 4-person team (analyst, scriptwriter, VO artist, video editor) grinding for 2-4 hours just to output 1-2 videos a day. This bottleneck means publishers are losing crucial audience share to fintech influencers who pump out lower-quality editorial content faster. We wanted to completely automate this process and capture that "golden hour" right after the market closes.

What it does

We built Jarvis, an end-to-end autonomous market video engine that requires zero human editing. Jarvis scales production from a measly 1-2 videos a day to 50+. It pulls real-time financial data, writes the script, generates the voiceover, and renders a highly engaging, broadcast-ready vertical video in roughly 5 minutes. The best part? The cost per video drops from ~₹80K down to around ₹300.

How we built it

Our technical architecture is driven by a Multi-Agent Gemini Intelligence layer. We chained three specialized agents in sequence:

  • Analyst Agent: Cross-references NSE/BSE prices and RSS news to output market mood and contrarian insights.

  • Scriptwriter Agent: Breaks the analysis into 4-7 scenes tailored for a snappy 60-90 second format.

  • Director Agent: Assigns visual directives and word-index cues for the renderer.

To keep the LLM outputs deterministic and production-safe, we rigidly enforced schema validation using Pydantic. For the audio, we hooked up ElevenLabs for exact word-sync timestamps. Finally, the entire visual layer is built on React 18 and D3.js, pulling assets from Pexels, and rendered programmatically into an MP4 using Remotion 4.x and FFmpeg.

Challenges we ran into

Relying purely on LLMs for a fully autonomous pipeline is risky because of unpredictable outputs. Our biggest headache was ensuring fault tolerance and reliability. We implemented exponential backoff retries (3 attempts per agent). If Gemini completely fails, our Director agent gracefully degrades to a keyword-scan fallback so there is zero downtime. Creating a "Blueprint deterministic planner" ensured that a video always generates regardless of LLM quirks. Getting exact word-sync audio timestamps to perfectly trigger React animations (like candlestick charts or bar races) was also incredibly tough to dial in.

Accomplishments that we're proud of

  • Speed & Scale: We achieved a 97% faster delivery time, dropping the turnaround from 3 hours to just 5 minutes. Throughput increased 25x.

  • Dynamic Storytelling: We engineered 8 dynamic narrative arcs (like panic_calmer, fii_dii_tug, and global_shock). The system dynamically selects these based on real-time triggers, like a Nifty move of ±2% or VIX fluctuations, ensuring there are no "dead" content days.

  • Massive Cost Reduction: We reduced production costs by 96% while maintaining broadcast quality.

What we learned

We learned that structure is everything when dealing with generative AI. Using strict JSON schemas is the only way to make LLMs predictable enough for automated video rendering. We also realized the sheer business value of micro-automations. The economics of replacing a full studio team with API calls blew our minds. We mapped out our projected return using a standard payback evaluation:

$$ROI = \left( \frac{\text{Monthly Savings} - \text{Monthly API Cost}}{\text{Monthly API Cost}} \right) \times 100$$

With API costs running around ₹90K/month against traditional team savings of 12-15L/month, Jarvis literally pays for itself in less than a month, yielding an ROI of 1,500%+ in Year 1.

What's next for Jarvis

Now that we've nailed the publisher use-case, we want to expand into new strategic deployments. We plan to white-label Jarvis for Brokerages and AMCs so they can auto-generate personalized portfolio recap videos for their clients, which should boost retention by 25%. We also want to adapt the engine to pump out automated investor education content to help activate new users faster.

Built With

Share this project:

Updates

Submission history